1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <linux/kernel.h>
3 #include <linux/errno.h>
4 #include <linux/err.h>
5 #include <linux/spinlock.h>
6
7 #include <linux/mm.h>
8 #include <linux/memremap.h>
9 #include <linux/pagemap.h>
10 #include <linux/rmap.h>
11 #include <linux/swap.h>
12 #include <linux/swapops.h>
13 #include <linux/secretmem.h>
14
15 #include <linux/sched/signal.h>
16 #include <linux/rwsem.h>
17 #include <linux/hugetlb.h>
18 #include <linux/migrate.h>
19 #include <linux/mm_inline.h>
20 #include <linux/sched/mm.h>
21 #include <linux/shmem_fs.h>
22
23 #include <asm/mmu_context.h>
24 #include <asm/tlbflush.h>
25
26 #include "internal.h"
27
28 struct follow_page_context {
29 struct dev_pagemap *pgmap;
30 unsigned int page_mask;
31 };
32
sanity_check_pinned_pages(struct page ** pages,unsigned long npages)33 static inline void sanity_check_pinned_pages(struct page **pages,
34 unsigned long npages)
35 {
36 if (!IS_ENABLED(CONFIG_DEBUG_VM))
37 return;
38
39 /*
40 * We only pin anonymous pages if they are exclusive. Once pinned, we
41 * can no longer turn them possibly shared and PageAnonExclusive() will
42 * stick around until the page is freed.
43 *
44 * We'd like to verify that our pinned anonymous pages are still mapped
45 * exclusively. The issue with anon THP is that we don't know how
46 * they are/were mapped when pinning them. However, for anon
47 * THP we can assume that either the given page (PTE-mapped THP) or
48 * the head page (PMD-mapped THP) should be PageAnonExclusive(). If
49 * neither is the case, there is certainly something wrong.
50 */
51 for (; npages; npages--, pages++) {
52 struct page *page = *pages;
53 struct folio *folio = page_folio(page);
54
55 if (is_zero_page(page) ||
56 !folio_test_anon(folio))
57 continue;
58 if (!folio_test_large(folio) || folio_test_hugetlb(folio))
59 VM_BUG_ON_PAGE(!PageAnonExclusive(&folio->page), page);
60 else
61 /* Either a PTE-mapped or a PMD-mapped THP. */
62 VM_BUG_ON_PAGE(!PageAnonExclusive(&folio->page) &&
63 !PageAnonExclusive(page), page);
64 }
65 }
66
67 /*
68 * Return the folio with ref appropriately incremented,
69 * or NULL if that failed.
70 */
try_get_folio(struct page * page,int refs)71 static inline struct folio *try_get_folio(struct page *page, int refs)
72 {
73 struct folio *folio;
74
75 retry:
76 folio = page_folio(page);
77 if (WARN_ON_ONCE(folio_ref_count(folio) < 0))
78 return NULL;
79 if (unlikely(!folio_ref_try_add_rcu(folio, refs)))
80 return NULL;
81
82 /*
83 * At this point we have a stable reference to the folio; but it
84 * could be that between calling page_folio() and the refcount
85 * increment, the folio was split, in which case we'd end up
86 * holding a reference on a folio that has nothing to do with the page
87 * we were given anymore.
88 * So now that the folio is stable, recheck that the page still
89 * belongs to this folio.
90 */
91 if (unlikely(page_folio(page) != folio)) {
92 if (!put_devmap_managed_page_refs(&folio->page, refs))
93 folio_put_refs(folio, refs);
94 goto retry;
95 }
96
97 return folio;
98 }
99
100 /**
101 * try_grab_folio() - Attempt to get or pin a folio.
102 * @page: pointer to page to be grabbed
103 * @refs: the value to (effectively) add to the folio's refcount
104 * @flags: gup flags: these are the FOLL_* flag values.
105 *
106 * "grab" names in this file mean, "look at flags to decide whether to use
107 * FOLL_PIN or FOLL_GET behavior, when incrementing the folio's refcount.
108 *
109 * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the
110 * same time. (That's true throughout the get_user_pages*() and
111 * pin_user_pages*() APIs.) Cases:
112 *
113 * FOLL_GET: folio's refcount will be incremented by @refs.
114 *
115 * FOLL_PIN on large folios: folio's refcount will be incremented by
116 * @refs, and its pincount will be incremented by @refs.
117 *
118 * FOLL_PIN on single-page folios: folio's refcount will be incremented by
119 * @refs * GUP_PIN_COUNTING_BIAS.
120 *
121 * Return: The folio containing @page (with refcount appropriately
122 * incremented) for success, or NULL upon failure. If neither FOLL_GET
123 * nor FOLL_PIN was set, that's considered failure, and furthermore,
124 * a likely bug in the caller, so a warning is also emitted.
125 */
try_grab_folio(struct page * page,int refs,unsigned int flags)126 struct folio *try_grab_folio(struct page *page, int refs, unsigned int flags)
127 {
128 struct folio *folio;
129
130 if (WARN_ON_ONCE((flags & (FOLL_GET | FOLL_PIN)) == 0))
131 return NULL;
132
133 if (unlikely(!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)))
134 return NULL;
135
136 if (flags & FOLL_GET)
137 return try_get_folio(page, refs);
138
139 /* FOLL_PIN is set */
140
141 /*
142 * Don't take a pin on the zero page - it's not going anywhere
143 * and it is used in a *lot* of places.
144 */
145 if (is_zero_page(page))
146 return page_folio(page);
147
148 folio = try_get_folio(page, refs);
149 if (!folio)
150 return NULL;
151
152 /*
153 * Can't do FOLL_LONGTERM + FOLL_PIN gup fast path if not in a
154 * right zone, so fail and let the caller fall back to the slow
155 * path.
156 */
157 if (unlikely((flags & FOLL_LONGTERM) &&
158 !folio_is_longterm_pinnable(folio))) {
159 if (!put_devmap_managed_page_refs(&folio->page, refs))
160 folio_put_refs(folio, refs);
161 return NULL;
162 }
163
164 /*
165 * When pinning a large folio, use an exact count to track it.
166 *
167 * However, be sure to *also* increment the normal folio
168 * refcount field at least once, so that the folio really
169 * is pinned. That's why the refcount from the earlier
170 * try_get_folio() is left intact.
171 */
172 if (folio_test_large(folio))
173 atomic_add(refs, &folio->_pincount);
174 else
175 folio_ref_add(folio,
176 refs * (GUP_PIN_COUNTING_BIAS - 1));
177 /*
178 * Adjust the pincount before re-checking the PTE for changes.
179 * This is essentially a smp_mb() and is paired with a memory
180 * barrier in page_try_share_anon_rmap().
181 */
182 smp_mb__after_atomic();
183
184 node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, refs);
185
186 return folio;
187 }
188
gup_put_folio(struct folio * folio,int refs,unsigned int flags)189 static void gup_put_folio(struct folio *folio, int refs, unsigned int flags)
190 {
191 if (flags & FOLL_PIN) {
192 if (is_zero_folio(folio))
193 return;
194 node_stat_mod_folio(folio, NR_FOLL_PIN_RELEASED, refs);
195 if (folio_test_large(folio))
196 atomic_sub(refs, &folio->_pincount);
197 else
198 refs *= GUP_PIN_COUNTING_BIAS;
199 }
200
201 if (!put_devmap_managed_page_refs(&folio->page, refs))
202 folio_put_refs(folio, refs);
203 }
204
205 /**
206 * try_grab_page() - elevate a page's refcount by a flag-dependent amount
207 * @page: pointer to page to be grabbed
208 * @flags: gup flags: these are the FOLL_* flag values.
209 *
210 * This might not do anything at all, depending on the flags argument.
211 *
212 * "grab" names in this file mean, "look at flags to decide whether to use
213 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
214 *
215 * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
216 * time. Cases: please see the try_grab_folio() documentation, with
217 * "refs=1".
218 *
219 * Return: 0 for success, or if no action was required (if neither FOLL_PIN
220 * nor FOLL_GET was set, nothing is done). A negative error code for failure:
221 *
222 * -ENOMEM FOLL_GET or FOLL_PIN was set, but the page could not
223 * be grabbed.
224 */
try_grab_page(struct page * page,unsigned int flags)225 int __must_check try_grab_page(struct page *page, unsigned int flags)
226 {
227 struct folio *folio = page_folio(page);
228
229 if (WARN_ON_ONCE(folio_ref_count(folio) <= 0))
230 return -ENOMEM;
231
232 if (unlikely(!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)))
233 return -EREMOTEIO;
234
235 if (flags & FOLL_GET)
236 folio_ref_inc(folio);
237 else if (flags & FOLL_PIN) {
238 /*
239 * Don't take a pin on the zero page - it's not going anywhere
240 * and it is used in a *lot* of places.
241 */
242 if (is_zero_page(page))
243 return 0;
244
245 /*
246 * Similar to try_grab_folio(): be sure to *also*
247 * increment the normal page refcount field at least once,
248 * so that the page really is pinned.
249 */
250 if (folio_test_large(folio)) {
251 folio_ref_add(folio, 1);
252 atomic_add(1, &folio->_pincount);
253 } else {
254 folio_ref_add(folio, GUP_PIN_COUNTING_BIAS);
255 }
256
257 node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, 1);
258 }
259
260 return 0;
261 }
262
263 /**
264 * unpin_user_page() - release a dma-pinned page
265 * @page: pointer to page to be released
266 *
267 * Pages that were pinned via pin_user_pages*() must be released via either
268 * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
269 * that such pages can be separately tracked and uniquely handled. In
270 * particular, interactions with RDMA and filesystems need special handling.
271 */
unpin_user_page(struct page * page)272 void unpin_user_page(struct page *page)
273 {
274 sanity_check_pinned_pages(&page, 1);
275 gup_put_folio(page_folio(page), 1, FOLL_PIN);
276 }
277 EXPORT_SYMBOL(unpin_user_page);
278
279 /**
280 * folio_add_pin - Try to get an additional pin on a pinned folio
281 * @folio: The folio to be pinned
282 *
283 * Get an additional pin on a folio we already have a pin on. Makes no change
284 * if the folio is a zero_page.
285 */
folio_add_pin(struct folio * folio)286 void folio_add_pin(struct folio *folio)
287 {
288 if (is_zero_folio(folio))
289 return;
290
291 /*
292 * Similar to try_grab_folio(): be sure to *also* increment the normal
293 * page refcount field at least once, so that the page really is
294 * pinned.
295 */
296 if (folio_test_large(folio)) {
297 WARN_ON_ONCE(atomic_read(&folio->_pincount) < 1);
298 folio_ref_inc(folio);
299 atomic_inc(&folio->_pincount);
300 } else {
301 WARN_ON_ONCE(folio_ref_count(folio) < GUP_PIN_COUNTING_BIAS);
302 folio_ref_add(folio, GUP_PIN_COUNTING_BIAS);
303 }
304 }
305
gup_folio_range_next(struct page * start,unsigned long npages,unsigned long i,unsigned int * ntails)306 static inline struct folio *gup_folio_range_next(struct page *start,
307 unsigned long npages, unsigned long i, unsigned int *ntails)
308 {
309 struct page *next = nth_page(start, i);
310 struct folio *folio = page_folio(next);
311 unsigned int nr = 1;
312
313 if (folio_test_large(folio))
314 nr = min_t(unsigned int, npages - i,
315 folio_nr_pages(folio) - folio_page_idx(folio, next));
316
317 *ntails = nr;
318 return folio;
319 }
320
gup_folio_next(struct page ** list,unsigned long npages,unsigned long i,unsigned int * ntails)321 static inline struct folio *gup_folio_next(struct page **list,
322 unsigned long npages, unsigned long i, unsigned int *ntails)
323 {
324 struct folio *folio = page_folio(list[i]);
325 unsigned int nr;
326
327 for (nr = i + 1; nr < npages; nr++) {
328 if (page_folio(list[nr]) != folio)
329 break;
330 }
331
332 *ntails = nr - i;
333 return folio;
334 }
335
336 /**
337 * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
338 * @pages: array of pages to be maybe marked dirty, and definitely released.
339 * @npages: number of pages in the @pages array.
340 * @make_dirty: whether to mark the pages dirty
341 *
342 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
343 * variants called on that page.
344 *
345 * For each page in the @pages array, make that page (or its head page, if a
346 * compound page) dirty, if @make_dirty is true, and if the page was previously
347 * listed as clean. In any case, releases all pages using unpin_user_page(),
348 * possibly via unpin_user_pages(), for the non-dirty case.
349 *
350 * Please see the unpin_user_page() documentation for details.
351 *
352 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
353 * required, then the caller should a) verify that this is really correct,
354 * because _lock() is usually required, and b) hand code it:
355 * set_page_dirty_lock(), unpin_user_page().
356 *
357 */
unpin_user_pages_dirty_lock(struct page ** pages,unsigned long npages,bool make_dirty)358 void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npages,
359 bool make_dirty)
360 {
361 unsigned long i;
362 struct folio *folio;
363 unsigned int nr;
364
365 if (!make_dirty) {
366 unpin_user_pages(pages, npages);
367 return;
368 }
369
370 sanity_check_pinned_pages(pages, npages);
371 for (i = 0; i < npages; i += nr) {
372 folio = gup_folio_next(pages, npages, i, &nr);
373 /*
374 * Checking PageDirty at this point may race with
375 * clear_page_dirty_for_io(), but that's OK. Two key
376 * cases:
377 *
378 * 1) This code sees the page as already dirty, so it
379 * skips the call to set_page_dirty(). That could happen
380 * because clear_page_dirty_for_io() called
381 * page_mkclean(), followed by set_page_dirty().
382 * However, now the page is going to get written back,
383 * which meets the original intention of setting it
384 * dirty, so all is well: clear_page_dirty_for_io() goes
385 * on to call TestClearPageDirty(), and write the page
386 * back.
387 *
388 * 2) This code sees the page as clean, so it calls
389 * set_page_dirty(). The page stays dirty, despite being
390 * written back, so it gets written back again in the
391 * next writeback cycle. This is harmless.
392 */
393 if (!folio_test_dirty(folio)) {
394 folio_lock(folio);
395 folio_mark_dirty(folio);
396 folio_unlock(folio);
397 }
398 gup_put_folio(folio, nr, FOLL_PIN);
399 }
400 }
401 EXPORT_SYMBOL(unpin_user_pages_dirty_lock);
402
403 /**
404 * unpin_user_page_range_dirty_lock() - release and optionally dirty
405 * gup-pinned page range
406 *
407 * @page: the starting page of a range maybe marked dirty, and definitely released.
408 * @npages: number of consecutive pages to release.
409 * @make_dirty: whether to mark the pages dirty
410 *
411 * "gup-pinned page range" refers to a range of pages that has had one of the
412 * pin_user_pages() variants called on that page.
413 *
414 * For the page ranges defined by [page .. page+npages], make that range (or
415 * its head pages, if a compound page) dirty, if @make_dirty is true, and if the
416 * page range was previously listed as clean.
417 *
418 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
419 * required, then the caller should a) verify that this is really correct,
420 * because _lock() is usually required, and b) hand code it:
421 * set_page_dirty_lock(), unpin_user_page().
422 *
423 */
unpin_user_page_range_dirty_lock(struct page * page,unsigned long npages,bool make_dirty)424 void unpin_user_page_range_dirty_lock(struct page *page, unsigned long npages,
425 bool make_dirty)
426 {
427 unsigned long i;
428 struct folio *folio;
429 unsigned int nr;
430
431 for (i = 0; i < npages; i += nr) {
432 folio = gup_folio_range_next(page, npages, i, &nr);
433 if (make_dirty && !folio_test_dirty(folio)) {
434 folio_lock(folio);
435 folio_mark_dirty(folio);
436 folio_unlock(folio);
437 }
438 gup_put_folio(folio, nr, FOLL_PIN);
439 }
440 }
441 EXPORT_SYMBOL(unpin_user_page_range_dirty_lock);
442
unpin_user_pages_lockless(struct page ** pages,unsigned long npages)443 static void unpin_user_pages_lockless(struct page **pages, unsigned long npages)
444 {
445 unsigned long i;
446 struct folio *folio;
447 unsigned int nr;
448
449 /*
450 * Don't perform any sanity checks because we might have raced with
451 * fork() and some anonymous pages might now actually be shared --
452 * which is why we're unpinning after all.
453 */
454 for (i = 0; i < npages; i += nr) {
455 folio = gup_folio_next(pages, npages, i, &nr);
456 gup_put_folio(folio, nr, FOLL_PIN);
457 }
458 }
459
460 /**
461 * unpin_user_pages() - release an array of gup-pinned pages.
462 * @pages: array of pages to be marked dirty and released.
463 * @npages: number of pages in the @pages array.
464 *
465 * For each page in the @pages array, release the page using unpin_user_page().
466 *
467 * Please see the unpin_user_page() documentation for details.
468 */
unpin_user_pages(struct page ** pages,unsigned long npages)469 void unpin_user_pages(struct page **pages, unsigned long npages)
470 {
471 unsigned long i;
472 struct folio *folio;
473 unsigned int nr;
474
475 /*
476 * If this WARN_ON() fires, then the system *might* be leaking pages (by
477 * leaving them pinned), but probably not. More likely, gup/pup returned
478 * a hard -ERRNO error to the caller, who erroneously passed it here.
479 */
480 if (WARN_ON(IS_ERR_VALUE(npages)))
481 return;
482
483 sanity_check_pinned_pages(pages, npages);
484 for (i = 0; i < npages; i += nr) {
485 folio = gup_folio_next(pages, npages, i, &nr);
486 gup_put_folio(folio, nr, FOLL_PIN);
487 }
488 }
489 EXPORT_SYMBOL(unpin_user_pages);
490
491 /*
492 * Set the MMF_HAS_PINNED if not set yet; after set it'll be there for the mm's
493 * lifecycle. Avoid setting the bit unless necessary, or it might cause write
494 * cache bouncing on large SMP machines for concurrent pinned gups.
495 */
mm_set_has_pinned_flag(unsigned long * mm_flags)496 static inline void mm_set_has_pinned_flag(unsigned long *mm_flags)
497 {
498 if (!test_bit(MMF_HAS_PINNED, mm_flags))
499 set_bit(MMF_HAS_PINNED, mm_flags);
500 }
501
502 #ifdef CONFIG_MMU
no_page_table(struct vm_area_struct * vma,unsigned int flags)503 static struct page *no_page_table(struct vm_area_struct *vma,
504 unsigned int flags)
505 {
506 /*
507 * When core dumping an enormous anonymous area that nobody
508 * has touched so far, we don't want to allocate unnecessary pages or
509 * page tables. Return error instead of NULL to skip handle_mm_fault,
510 * then get_dump_page() will return NULL to leave a hole in the dump.
511 * But we can only make this optimization where a hole would surely
512 * be zero-filled if handle_mm_fault() actually did handle it.
513 */
514 if ((flags & FOLL_DUMP) &&
515 (vma_is_anonymous(vma) || !vma->vm_ops->fault))
516 return ERR_PTR(-EFAULT);
517 return NULL;
518 }
519
follow_pfn_pte(struct vm_area_struct * vma,unsigned long address,pte_t * pte,unsigned int flags)520 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
521 pte_t *pte, unsigned int flags)
522 {
523 if (flags & FOLL_TOUCH) {
524 pte_t orig_entry = ptep_get(pte);
525 pte_t entry = orig_entry;
526
527 if (flags & FOLL_WRITE)
528 entry = pte_mkdirty(entry);
529 entry = pte_mkyoung(entry);
530
531 if (!pte_same(orig_entry, entry)) {
532 set_pte_at(vma->vm_mm, address, pte, entry);
533 update_mmu_cache(vma, address, pte);
534 }
535 }
536
537 /* Proper page table entry exists, but no corresponding struct page */
538 return -EEXIST;
539 }
540
541 /* FOLL_FORCE can write to even unwritable PTEs in COW mappings. */
can_follow_write_pte(pte_t pte,struct page * page,struct vm_area_struct * vma,unsigned int flags)542 static inline bool can_follow_write_pte(pte_t pte, struct page *page,
543 struct vm_area_struct *vma,
544 unsigned int flags)
545 {
546 /* If the pte is writable, we can write to the page. */
547 if (pte_write(pte))
548 return true;
549
550 /* Maybe FOLL_FORCE is set to override it? */
551 if (!(flags & FOLL_FORCE))
552 return false;
553
554 /* But FOLL_FORCE has no effect on shared mappings */
555 if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED))
556 return false;
557
558 /* ... or read-only private ones */
559 if (!(vma->vm_flags & VM_MAYWRITE))
560 return false;
561
562 /* ... or already writable ones that just need to take a write fault */
563 if (vma->vm_flags & VM_WRITE)
564 return false;
565
566 /*
567 * See can_change_pte_writable(): we broke COW and could map the page
568 * writable if we have an exclusive anonymous page ...
569 */
570 if (!page || !PageAnon(page) || !PageAnonExclusive(page))
571 return false;
572
573 /* ... and a write-fault isn't required for other reasons. */
574 if (vma_soft_dirty_enabled(vma) && !pte_soft_dirty(pte))
575 return false;
576 return !userfaultfd_pte_wp(vma, pte);
577 }
578
follow_page_pte(struct vm_area_struct * vma,unsigned long address,pmd_t * pmd,unsigned int flags,struct dev_pagemap ** pgmap)579 static struct page *follow_page_pte(struct vm_area_struct *vma,
580 unsigned long address, pmd_t *pmd, unsigned int flags,
581 struct dev_pagemap **pgmap)
582 {
583 struct mm_struct *mm = vma->vm_mm;
584 struct page *page;
585 spinlock_t *ptl;
586 pte_t *ptep, pte;
587 int ret;
588
589 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
590 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
591 (FOLL_PIN | FOLL_GET)))
592 return ERR_PTR(-EINVAL);
593
594 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
595 if (!ptep)
596 return no_page_table(vma, flags);
597 pte = ptep_get(ptep);
598 if (!pte_present(pte))
599 goto no_page;
600 if (pte_protnone(pte) && !gup_can_follow_protnone(vma, flags))
601 goto no_page;
602
603 page = vm_normal_page(vma, address, pte);
604
605 /*
606 * We only care about anon pages in can_follow_write_pte() and don't
607 * have to worry about pte_devmap() because they are never anon.
608 */
609 if ((flags & FOLL_WRITE) &&
610 !can_follow_write_pte(pte, page, vma, flags)) {
611 page = NULL;
612 goto out;
613 }
614
615 if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) {
616 /*
617 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
618 * case since they are only valid while holding the pgmap
619 * reference.
620 */
621 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
622 if (*pgmap)
623 page = pte_page(pte);
624 else
625 goto no_page;
626 } else if (unlikely(!page)) {
627 if (flags & FOLL_DUMP) {
628 /* Avoid special (like zero) pages in core dumps */
629 page = ERR_PTR(-EFAULT);
630 goto out;
631 }
632
633 if (is_zero_pfn(pte_pfn(pte))) {
634 page = pte_page(pte);
635 } else {
636 ret = follow_pfn_pte(vma, address, ptep, flags);
637 page = ERR_PTR(ret);
638 goto out;
639 }
640 }
641
642 if (!pte_write(pte) && gup_must_unshare(vma, flags, page)) {
643 page = ERR_PTR(-EMLINK);
644 goto out;
645 }
646
647 VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
648 !PageAnonExclusive(page), page);
649
650 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
651 ret = try_grab_page(page, flags);
652 if (unlikely(ret)) {
653 page = ERR_PTR(ret);
654 goto out;
655 }
656
657 /*
658 * We need to make the page accessible if and only if we are going
659 * to access its content (the FOLL_PIN case). Please see
660 * Documentation/core-api/pin_user_pages.rst for details.
661 */
662 if (flags & FOLL_PIN) {
663 ret = arch_make_page_accessible(page);
664 if (ret) {
665 unpin_user_page(page);
666 page = ERR_PTR(ret);
667 goto out;
668 }
669 }
670 if (flags & FOLL_TOUCH) {
671 if ((flags & FOLL_WRITE) &&
672 !pte_dirty(pte) && !PageDirty(page))
673 set_page_dirty(page);
674 /*
675 * pte_mkyoung() would be more correct here, but atomic care
676 * is needed to avoid losing the dirty bit: it is easier to use
677 * mark_page_accessed().
678 */
679 mark_page_accessed(page);
680 }
681 out:
682 pte_unmap_unlock(ptep, ptl);
683 return page;
684 no_page:
685 pte_unmap_unlock(ptep, ptl);
686 if (!pte_none(pte))
687 return NULL;
688 return no_page_table(vma, flags);
689 }
690
follow_pmd_mask(struct vm_area_struct * vma,unsigned long address,pud_t * pudp,unsigned int flags,struct follow_page_context * ctx)691 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
692 unsigned long address, pud_t *pudp,
693 unsigned int flags,
694 struct follow_page_context *ctx)
695 {
696 pmd_t *pmd, pmdval;
697 spinlock_t *ptl;
698 struct page *page;
699 struct mm_struct *mm = vma->vm_mm;
700
701 pmd = pmd_offset(pudp, address);
702 pmdval = pmdp_get_lockless(pmd);
703 if (pmd_none(pmdval))
704 return no_page_table(vma, flags);
705 if (!pmd_present(pmdval))
706 return no_page_table(vma, flags);
707 if (pmd_devmap(pmdval)) {
708 ptl = pmd_lock(mm, pmd);
709 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
710 spin_unlock(ptl);
711 if (page)
712 return page;
713 }
714 if (likely(!pmd_trans_huge(pmdval)))
715 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
716
717 if (pmd_protnone(pmdval) && !gup_can_follow_protnone(vma, flags))
718 return no_page_table(vma, flags);
719
720 ptl = pmd_lock(mm, pmd);
721 if (unlikely(!pmd_present(*pmd))) {
722 spin_unlock(ptl);
723 return no_page_table(vma, flags);
724 }
725 if (unlikely(!pmd_trans_huge(*pmd))) {
726 spin_unlock(ptl);
727 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
728 }
729 if (flags & FOLL_SPLIT_PMD) {
730 spin_unlock(ptl);
731 split_huge_pmd(vma, pmd, address);
732 /* If pmd was left empty, stuff a page table in there quickly */
733 return pte_alloc(mm, pmd) ? ERR_PTR(-ENOMEM) :
734 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
735 }
736 page = follow_trans_huge_pmd(vma, address, pmd, flags);
737 spin_unlock(ptl);
738 ctx->page_mask = HPAGE_PMD_NR - 1;
739 return page;
740 }
741
follow_pud_mask(struct vm_area_struct * vma,unsigned long address,p4d_t * p4dp,unsigned int flags,struct follow_page_context * ctx)742 static struct page *follow_pud_mask(struct vm_area_struct *vma,
743 unsigned long address, p4d_t *p4dp,
744 unsigned int flags,
745 struct follow_page_context *ctx)
746 {
747 pud_t *pud;
748 spinlock_t *ptl;
749 struct page *page;
750 struct mm_struct *mm = vma->vm_mm;
751
752 pud = pud_offset(p4dp, address);
753 if (pud_none(*pud))
754 return no_page_table(vma, flags);
755 if (pud_devmap(*pud)) {
756 ptl = pud_lock(mm, pud);
757 page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
758 spin_unlock(ptl);
759 if (page)
760 return page;
761 }
762 if (unlikely(pud_bad(*pud)))
763 return no_page_table(vma, flags);
764
765 return follow_pmd_mask(vma, address, pud, flags, ctx);
766 }
767
follow_p4d_mask(struct vm_area_struct * vma,unsigned long address,pgd_t * pgdp,unsigned int flags,struct follow_page_context * ctx)768 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
769 unsigned long address, pgd_t *pgdp,
770 unsigned int flags,
771 struct follow_page_context *ctx)
772 {
773 p4d_t *p4d;
774
775 p4d = p4d_offset(pgdp, address);
776 if (p4d_none(*p4d))
777 return no_page_table(vma, flags);
778 BUILD_BUG_ON(p4d_huge(*p4d));
779 if (unlikely(p4d_bad(*p4d)))
780 return no_page_table(vma, flags);
781
782 return follow_pud_mask(vma, address, p4d, flags, ctx);
783 }
784
785 /**
786 * follow_page_mask - look up a page descriptor from a user-virtual address
787 * @vma: vm_area_struct mapping @address
788 * @address: virtual address to look up
789 * @flags: flags modifying lookup behaviour
790 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
791 * pointer to output page_mask
792 *
793 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
794 *
795 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
796 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
797 *
798 * When getting an anonymous page and the caller has to trigger unsharing
799 * of a shared anonymous page first, -EMLINK is returned. The caller should
800 * trigger a fault with FAULT_FLAG_UNSHARE set. Note that unsharing is only
801 * relevant with FOLL_PIN and !FOLL_WRITE.
802 *
803 * On output, the @ctx->page_mask is set according to the size of the page.
804 *
805 * Return: the mapped (struct page *), %NULL if no mapping exists, or
806 * an error pointer if there is a mapping to something not represented
807 * by a page descriptor (see also vm_normal_page()).
808 */
follow_page_mask(struct vm_area_struct * vma,unsigned long address,unsigned int flags,struct follow_page_context * ctx)809 static struct page *follow_page_mask(struct vm_area_struct *vma,
810 unsigned long address, unsigned int flags,
811 struct follow_page_context *ctx)
812 {
813 pgd_t *pgd;
814 struct mm_struct *mm = vma->vm_mm;
815
816 ctx->page_mask = 0;
817
818 /*
819 * Call hugetlb_follow_page_mask for hugetlb vmas as it will use
820 * special hugetlb page table walking code. This eliminates the
821 * need to check for hugetlb entries in the general walking code.
822 */
823 if (is_vm_hugetlb_page(vma))
824 return hugetlb_follow_page_mask(vma, address, flags,
825 &ctx->page_mask);
826
827 pgd = pgd_offset(mm, address);
828
829 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
830 return no_page_table(vma, flags);
831
832 return follow_p4d_mask(vma, address, pgd, flags, ctx);
833 }
834
follow_page(struct vm_area_struct * vma,unsigned long address,unsigned int foll_flags)835 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
836 unsigned int foll_flags)
837 {
838 struct follow_page_context ctx = { NULL };
839 struct page *page;
840
841 if (vma_is_secretmem(vma))
842 return NULL;
843
844 if (WARN_ON_ONCE(foll_flags & FOLL_PIN))
845 return NULL;
846
847 /*
848 * We never set FOLL_HONOR_NUMA_FAULT because callers don't expect
849 * to fail on PROT_NONE-mapped pages.
850 */
851 page = follow_page_mask(vma, address, foll_flags, &ctx);
852 if (ctx.pgmap)
853 put_dev_pagemap(ctx.pgmap);
854 return page;
855 }
856
get_gate_page(struct mm_struct * mm,unsigned long address,unsigned int gup_flags,struct vm_area_struct ** vma,struct page ** page)857 static int get_gate_page(struct mm_struct *mm, unsigned long address,
858 unsigned int gup_flags, struct vm_area_struct **vma,
859 struct page **page)
860 {
861 pgd_t *pgd;
862 p4d_t *p4d;
863 pud_t *pud;
864 pmd_t *pmd;
865 pte_t *pte;
866 pte_t entry;
867 int ret = -EFAULT;
868
869 /* user gate pages are read-only */
870 if (gup_flags & FOLL_WRITE)
871 return -EFAULT;
872 if (address > TASK_SIZE)
873 pgd = pgd_offset_k(address);
874 else
875 pgd = pgd_offset_gate(mm, address);
876 if (pgd_none(*pgd))
877 return -EFAULT;
878 p4d = p4d_offset(pgd, address);
879 if (p4d_none(*p4d))
880 return -EFAULT;
881 pud = pud_offset(p4d, address);
882 if (pud_none(*pud))
883 return -EFAULT;
884 pmd = pmd_offset(pud, address);
885 if (!pmd_present(*pmd))
886 return -EFAULT;
887 pte = pte_offset_map(pmd, address);
888 if (!pte)
889 return -EFAULT;
890 entry = ptep_get(pte);
891 if (pte_none(entry))
892 goto unmap;
893 *vma = get_gate_vma(mm);
894 if (!page)
895 goto out;
896 *page = vm_normal_page(*vma, address, entry);
897 if (!*page) {
898 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(entry)))
899 goto unmap;
900 *page = pte_page(entry);
901 }
902 ret = try_grab_page(*page, gup_flags);
903 if (unlikely(ret))
904 goto unmap;
905 out:
906 ret = 0;
907 unmap:
908 pte_unmap(pte);
909 return ret;
910 }
911
912 /*
913 * mmap_lock must be held on entry. If @flags has FOLL_UNLOCKABLE but not
914 * FOLL_NOWAIT, the mmap_lock may be released. If it is, *@locked will be set
915 * to 0 and -EBUSY returned.
916 */
faultin_page(struct vm_area_struct * vma,unsigned long address,unsigned int * flags,bool unshare,int * locked)917 static int faultin_page(struct vm_area_struct *vma,
918 unsigned long address, unsigned int *flags, bool unshare,
919 int *locked)
920 {
921 unsigned int fault_flags = 0;
922 vm_fault_t ret;
923
924 if (*flags & FOLL_NOFAULT)
925 return -EFAULT;
926 if (*flags & FOLL_WRITE)
927 fault_flags |= FAULT_FLAG_WRITE;
928 if (*flags & FOLL_REMOTE)
929 fault_flags |= FAULT_FLAG_REMOTE;
930 if (*flags & FOLL_UNLOCKABLE) {
931 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
932 /*
933 * FAULT_FLAG_INTERRUPTIBLE is opt-in. GUP callers must set
934 * FOLL_INTERRUPTIBLE to enable FAULT_FLAG_INTERRUPTIBLE.
935 * That's because some callers may not be prepared to
936 * handle early exits caused by non-fatal signals.
937 */
938 if (*flags & FOLL_INTERRUPTIBLE)
939 fault_flags |= FAULT_FLAG_INTERRUPTIBLE;
940 }
941 if (*flags & FOLL_NOWAIT)
942 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
943 if (*flags & FOLL_TRIED) {
944 /*
945 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
946 * can co-exist
947 */
948 fault_flags |= FAULT_FLAG_TRIED;
949 }
950 if (unshare) {
951 fault_flags |= FAULT_FLAG_UNSHARE;
952 /* FAULT_FLAG_WRITE and FAULT_FLAG_UNSHARE are incompatible */
953 VM_BUG_ON(fault_flags & FAULT_FLAG_WRITE);
954 }
955
956 ret = handle_mm_fault(vma, address, fault_flags, NULL);
957
958 if (ret & VM_FAULT_COMPLETED) {
959 /*
960 * With FAULT_FLAG_RETRY_NOWAIT we'll never release the
961 * mmap lock in the page fault handler. Sanity check this.
962 */
963 WARN_ON_ONCE(fault_flags & FAULT_FLAG_RETRY_NOWAIT);
964 *locked = 0;
965
966 /*
967 * We should do the same as VM_FAULT_RETRY, but let's not
968 * return -EBUSY since that's not reflecting the reality of
969 * what has happened - we've just fully completed a page
970 * fault, with the mmap lock released. Use -EAGAIN to show
971 * that we want to take the mmap lock _again_.
972 */
973 return -EAGAIN;
974 }
975
976 if (ret & VM_FAULT_ERROR) {
977 int err = vm_fault_to_errno(ret, *flags);
978
979 if (err)
980 return err;
981 BUG();
982 }
983
984 if (ret & VM_FAULT_RETRY) {
985 if (!(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
986 *locked = 0;
987 return -EBUSY;
988 }
989
990 return 0;
991 }
992
993 /*
994 * Writing to file-backed mappings which require folio dirty tracking using GUP
995 * is a fundamentally broken operation, as kernel write access to GUP mappings
996 * do not adhere to the semantics expected by a file system.
997 *
998 * Consider the following scenario:-
999 *
1000 * 1. A folio is written to via GUP which write-faults the memory, notifying
1001 * the file system and dirtying the folio.
1002 * 2. Later, writeback is triggered, resulting in the folio being cleaned and
1003 * the PTE being marked read-only.
1004 * 3. The GUP caller writes to the folio, as it is mapped read/write via the
1005 * direct mapping.
1006 * 4. The GUP caller, now done with the page, unpins it and sets it dirty
1007 * (though it does not have to).
1008 *
1009 * This results in both data being written to a folio without writenotify, and
1010 * the folio being dirtied unexpectedly (if the caller decides to do so).
1011 */
writable_file_mapping_allowed(struct vm_area_struct * vma,unsigned long gup_flags)1012 static bool writable_file_mapping_allowed(struct vm_area_struct *vma,
1013 unsigned long gup_flags)
1014 {
1015 /*
1016 * If we aren't pinning then no problematic write can occur. A long term
1017 * pin is the most egregious case so this is the case we disallow.
1018 */
1019 if ((gup_flags & (FOLL_PIN | FOLL_LONGTERM)) !=
1020 (FOLL_PIN | FOLL_LONGTERM))
1021 return true;
1022
1023 /*
1024 * If the VMA does not require dirty tracking then no problematic write
1025 * can occur either.
1026 */
1027 return !vma_needs_dirty_tracking(vma);
1028 }
1029
check_vma_flags(struct vm_area_struct * vma,unsigned long gup_flags)1030 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
1031 {
1032 vm_flags_t vm_flags = vma->vm_flags;
1033 int write = (gup_flags & FOLL_WRITE);
1034 int foreign = (gup_flags & FOLL_REMOTE);
1035 bool vma_anon = vma_is_anonymous(vma);
1036
1037 if (vm_flags & (VM_IO | VM_PFNMAP))
1038 return -EFAULT;
1039
1040 if ((gup_flags & FOLL_ANON) && !vma_anon)
1041 return -EFAULT;
1042
1043 if ((gup_flags & FOLL_LONGTERM) && vma_is_fsdax(vma))
1044 return -EOPNOTSUPP;
1045
1046 if (vma_is_secretmem(vma))
1047 return -EFAULT;
1048
1049 if (write) {
1050 if (!vma_anon &&
1051 !writable_file_mapping_allowed(vma, gup_flags))
1052 return -EFAULT;
1053
1054 if (!(vm_flags & VM_WRITE) || (vm_flags & VM_SHADOW_STACK)) {
1055 if (!(gup_flags & FOLL_FORCE))
1056 return -EFAULT;
1057 /* hugetlb does not support FOLL_FORCE|FOLL_WRITE. */
1058 if (is_vm_hugetlb_page(vma))
1059 return -EFAULT;
1060 /*
1061 * We used to let the write,force case do COW in a
1062 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
1063 * set a breakpoint in a read-only mapping of an
1064 * executable, without corrupting the file (yet only
1065 * when that file had been opened for writing!).
1066 * Anon pages in shared mappings are surprising: now
1067 * just reject it.
1068 */
1069 if (!is_cow_mapping(vm_flags))
1070 return -EFAULT;
1071 }
1072 } else if (!(vm_flags & VM_READ)) {
1073 if (!(gup_flags & FOLL_FORCE))
1074 return -EFAULT;
1075 /*
1076 * Is there actually any vma we can reach here which does not
1077 * have VM_MAYREAD set?
1078 */
1079 if (!(vm_flags & VM_MAYREAD))
1080 return -EFAULT;
1081 }
1082 /*
1083 * gups are always data accesses, not instruction
1084 * fetches, so execute=false here
1085 */
1086 if (!arch_vma_access_permitted(vma, write, false, foreign))
1087 return -EFAULT;
1088 return 0;
1089 }
1090
1091 /*
1092 * This is "vma_lookup()", but with a warning if we would have
1093 * historically expanded the stack in the GUP code.
1094 */
gup_vma_lookup(struct mm_struct * mm,unsigned long addr)1095 static struct vm_area_struct *gup_vma_lookup(struct mm_struct *mm,
1096 unsigned long addr)
1097 {
1098 #ifdef CONFIG_STACK_GROWSUP
1099 return vma_lookup(mm, addr);
1100 #else
1101 static volatile unsigned long next_warn;
1102 struct vm_area_struct *vma;
1103 unsigned long now, next;
1104
1105 vma = find_vma(mm, addr);
1106 if (!vma || (addr >= vma->vm_start))
1107 return vma;
1108
1109 /* Only warn for half-way relevant accesses */
1110 if (!(vma->vm_flags & VM_GROWSDOWN))
1111 return NULL;
1112 if (vma->vm_start - addr > 65536)
1113 return NULL;
1114
1115 /* Let's not warn more than once an hour.. */
1116 now = jiffies; next = next_warn;
1117 if (next && time_before(now, next))
1118 return NULL;
1119 next_warn = now + 60*60*HZ;
1120
1121 /* Let people know things may have changed. */
1122 pr_warn("GUP no longer grows the stack in %s (%d): %lx-%lx (%lx)\n",
1123 current->comm, task_pid_nr(current),
1124 vma->vm_start, vma->vm_end, addr);
1125 dump_stack();
1126 return NULL;
1127 #endif
1128 }
1129
1130 /**
1131 * __get_user_pages() - pin user pages in memory
1132 * @mm: mm_struct of target mm
1133 * @start: starting user address
1134 * @nr_pages: number of pages from start to pin
1135 * @gup_flags: flags modifying pin behaviour
1136 * @pages: array that receives pointers to the pages pinned.
1137 * Should be at least nr_pages long. Or NULL, if caller
1138 * only intends to ensure the pages are faulted in.
1139 * @locked: whether we're still with the mmap_lock held
1140 *
1141 * Returns either number of pages pinned (which may be less than the
1142 * number requested), or an error. Details about the return value:
1143 *
1144 * -- If nr_pages is 0, returns 0.
1145 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1146 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1147 * pages pinned. Again, this may be less than nr_pages.
1148 * -- 0 return value is possible when the fault would need to be retried.
1149 *
1150 * The caller is responsible for releasing returned @pages, via put_page().
1151 *
1152 * Must be called with mmap_lock held. It may be released. See below.
1153 *
1154 * __get_user_pages walks a process's page tables and takes a reference to
1155 * each struct page that each user address corresponds to at a given
1156 * instant. That is, it takes the page that would be accessed if a user
1157 * thread accesses the given user virtual address at that instant.
1158 *
1159 * This does not guarantee that the page exists in the user mappings when
1160 * __get_user_pages returns, and there may even be a completely different
1161 * page there in some cases (eg. if mmapped pagecache has been invalidated
1162 * and subsequently re-faulted). However it does guarantee that the page
1163 * won't be freed completely. And mostly callers simply care that the page
1164 * contains data that was valid *at some point in time*. Typically, an IO
1165 * or similar operation cannot guarantee anything stronger anyway because
1166 * locks can't be held over the syscall boundary.
1167 *
1168 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1169 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1170 * appropriate) must be called after the page is finished with, and
1171 * before put_page is called.
1172 *
1173 * If FOLL_UNLOCKABLE is set without FOLL_NOWAIT then the mmap_lock may
1174 * be released. If this happens *@locked will be set to 0 on return.
1175 *
1176 * A caller using such a combination of @gup_flags must therefore hold the
1177 * mmap_lock for reading only, and recognize when it's been released. Otherwise,
1178 * it must be held for either reading or writing and will not be released.
1179 *
1180 * In most cases, get_user_pages or get_user_pages_fast should be used
1181 * instead of __get_user_pages. __get_user_pages should be used only if
1182 * you need some special @gup_flags.
1183 */
__get_user_pages(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,int * locked)1184 static long __get_user_pages(struct mm_struct *mm,
1185 unsigned long start, unsigned long nr_pages,
1186 unsigned int gup_flags, struct page **pages,
1187 int *locked)
1188 {
1189 long ret = 0, i = 0;
1190 struct vm_area_struct *vma = NULL;
1191 struct follow_page_context ctx = { NULL };
1192
1193 if (!nr_pages)
1194 return 0;
1195
1196 start = untagged_addr_remote(mm, start);
1197
1198 VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN)));
1199
1200 do {
1201 struct page *page;
1202 unsigned int foll_flags = gup_flags;
1203 unsigned int page_increm;
1204
1205 /* first iteration or cross vma bound */
1206 if (!vma || start >= vma->vm_end) {
1207 /*
1208 * MADV_POPULATE_(READ|WRITE) wants to handle VMA
1209 * lookups+error reporting differently.
1210 */
1211 if (gup_flags & FOLL_MADV_POPULATE) {
1212 vma = vma_lookup(mm, start);
1213 if (!vma) {
1214 ret = -ENOMEM;
1215 goto out;
1216 }
1217 if (check_vma_flags(vma, gup_flags)) {
1218 ret = -EINVAL;
1219 goto out;
1220 }
1221 goto retry;
1222 }
1223 vma = gup_vma_lookup(mm, start);
1224 if (!vma && in_gate_area(mm, start)) {
1225 ret = get_gate_page(mm, start & PAGE_MASK,
1226 gup_flags, &vma,
1227 pages ? &page : NULL);
1228 if (ret)
1229 goto out;
1230 ctx.page_mask = 0;
1231 goto next_page;
1232 }
1233
1234 if (!vma) {
1235 ret = -EFAULT;
1236 goto out;
1237 }
1238 ret = check_vma_flags(vma, gup_flags);
1239 if (ret)
1240 goto out;
1241 }
1242 retry:
1243 /*
1244 * If we have a pending SIGKILL, don't keep faulting pages and
1245 * potentially allocating memory.
1246 */
1247 if (fatal_signal_pending(current)) {
1248 ret = -EINTR;
1249 goto out;
1250 }
1251 cond_resched();
1252
1253 page = follow_page_mask(vma, start, foll_flags, &ctx);
1254 if (!page || PTR_ERR(page) == -EMLINK) {
1255 ret = faultin_page(vma, start, &foll_flags,
1256 PTR_ERR(page) == -EMLINK, locked);
1257 switch (ret) {
1258 case 0:
1259 goto retry;
1260 case -EBUSY:
1261 case -EAGAIN:
1262 ret = 0;
1263 fallthrough;
1264 case -EFAULT:
1265 case -ENOMEM:
1266 case -EHWPOISON:
1267 goto out;
1268 }
1269 BUG();
1270 } else if (PTR_ERR(page) == -EEXIST) {
1271 /*
1272 * Proper page table entry exists, but no corresponding
1273 * struct page. If the caller expects **pages to be
1274 * filled in, bail out now, because that can't be done
1275 * for this page.
1276 */
1277 if (pages) {
1278 ret = PTR_ERR(page);
1279 goto out;
1280 }
1281 } else if (IS_ERR(page)) {
1282 ret = PTR_ERR(page);
1283 goto out;
1284 }
1285 next_page:
1286 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
1287 if (page_increm > nr_pages)
1288 page_increm = nr_pages;
1289
1290 if (pages) {
1291 struct page *subpage;
1292 unsigned int j;
1293
1294 /*
1295 * This must be a large folio (and doesn't need to
1296 * be the whole folio; it can be part of it), do
1297 * the refcount work for all the subpages too.
1298 *
1299 * NOTE: here the page may not be the head page
1300 * e.g. when start addr is not thp-size aligned.
1301 * try_grab_folio() should have taken care of tail
1302 * pages.
1303 */
1304 if (page_increm > 1) {
1305 struct folio *folio;
1306
1307 /*
1308 * Since we already hold refcount on the
1309 * large folio, this should never fail.
1310 */
1311 folio = try_grab_folio(page, page_increm - 1,
1312 foll_flags);
1313 if (WARN_ON_ONCE(!folio)) {
1314 /*
1315 * Release the 1st page ref if the
1316 * folio is problematic, fail hard.
1317 */
1318 gup_put_folio(page_folio(page), 1,
1319 foll_flags);
1320 ret = -EFAULT;
1321 goto out;
1322 }
1323 }
1324
1325 for (j = 0; j < page_increm; j++) {
1326 subpage = nth_page(page, j);
1327 pages[i + j] = subpage;
1328 flush_anon_page(vma, subpage, start + j * PAGE_SIZE);
1329 flush_dcache_page(subpage);
1330 }
1331 }
1332
1333 i += page_increm;
1334 start += page_increm * PAGE_SIZE;
1335 nr_pages -= page_increm;
1336 } while (nr_pages);
1337 out:
1338 if (ctx.pgmap)
1339 put_dev_pagemap(ctx.pgmap);
1340 return i ? i : ret;
1341 }
1342
vma_permits_fault(struct vm_area_struct * vma,unsigned int fault_flags)1343 static bool vma_permits_fault(struct vm_area_struct *vma,
1344 unsigned int fault_flags)
1345 {
1346 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
1347 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
1348 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
1349
1350 if (!(vm_flags & vma->vm_flags))
1351 return false;
1352
1353 /*
1354 * The architecture might have a hardware protection
1355 * mechanism other than read/write that can deny access.
1356 *
1357 * gup always represents data access, not instruction
1358 * fetches, so execute=false here:
1359 */
1360 if (!arch_vma_access_permitted(vma, write, false, foreign))
1361 return false;
1362
1363 return true;
1364 }
1365
1366 /**
1367 * fixup_user_fault() - manually resolve a user page fault
1368 * @mm: mm_struct of target mm
1369 * @address: user address
1370 * @fault_flags:flags to pass down to handle_mm_fault()
1371 * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller
1372 * does not allow retry. If NULL, the caller must guarantee
1373 * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1374 *
1375 * This is meant to be called in the specific scenario where for locking reasons
1376 * we try to access user memory in atomic context (within a pagefault_disable()
1377 * section), this returns -EFAULT, and we want to resolve the user fault before
1378 * trying again.
1379 *
1380 * Typically this is meant to be used by the futex code.
1381 *
1382 * The main difference with get_user_pages() is that this function will
1383 * unconditionally call handle_mm_fault() which will in turn perform all the
1384 * necessary SW fixup of the dirty and young bits in the PTE, while
1385 * get_user_pages() only guarantees to update these in the struct page.
1386 *
1387 * This is important for some architectures where those bits also gate the
1388 * access permission to the page because they are maintained in software. On
1389 * such architectures, gup() will not be enough to make a subsequent access
1390 * succeed.
1391 *
1392 * This function will not return with an unlocked mmap_lock. So it has not the
1393 * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1394 */
fixup_user_fault(struct mm_struct * mm,unsigned long address,unsigned int fault_flags,bool * unlocked)1395 int fixup_user_fault(struct mm_struct *mm,
1396 unsigned long address, unsigned int fault_flags,
1397 bool *unlocked)
1398 {
1399 struct vm_area_struct *vma;
1400 vm_fault_t ret;
1401
1402 address = untagged_addr_remote(mm, address);
1403
1404 if (unlocked)
1405 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1406
1407 retry:
1408 vma = gup_vma_lookup(mm, address);
1409 if (!vma)
1410 return -EFAULT;
1411
1412 if (!vma_permits_fault(vma, fault_flags))
1413 return -EFAULT;
1414
1415 if ((fault_flags & FAULT_FLAG_KILLABLE) &&
1416 fatal_signal_pending(current))
1417 return -EINTR;
1418
1419 ret = handle_mm_fault(vma, address, fault_flags, NULL);
1420
1421 if (ret & VM_FAULT_COMPLETED) {
1422 /*
1423 * NOTE: it's a pity that we need to retake the lock here
1424 * to pair with the unlock() in the callers. Ideally we
1425 * could tell the callers so they do not need to unlock.
1426 */
1427 mmap_read_lock(mm);
1428 *unlocked = true;
1429 return 0;
1430 }
1431
1432 if (ret & VM_FAULT_ERROR) {
1433 int err = vm_fault_to_errno(ret, 0);
1434
1435 if (err)
1436 return err;
1437 BUG();
1438 }
1439
1440 if (ret & VM_FAULT_RETRY) {
1441 mmap_read_lock(mm);
1442 *unlocked = true;
1443 fault_flags |= FAULT_FLAG_TRIED;
1444 goto retry;
1445 }
1446
1447 return 0;
1448 }
1449 EXPORT_SYMBOL_GPL(fixup_user_fault);
1450
1451 /*
1452 * GUP always responds to fatal signals. When FOLL_INTERRUPTIBLE is
1453 * specified, it'll also respond to generic signals. The caller of GUP
1454 * that has FOLL_INTERRUPTIBLE should take care of the GUP interruption.
1455 */
gup_signal_pending(unsigned int flags)1456 static bool gup_signal_pending(unsigned int flags)
1457 {
1458 if (fatal_signal_pending(current))
1459 return true;
1460
1461 if (!(flags & FOLL_INTERRUPTIBLE))
1462 return false;
1463
1464 return signal_pending(current);
1465 }
1466
1467 /*
1468 * Locking: (*locked == 1) means that the mmap_lock has already been acquired by
1469 * the caller. This function may drop the mmap_lock. If it does so, then it will
1470 * set (*locked = 0).
1471 *
1472 * (*locked == 0) means that the caller expects this function to acquire and
1473 * drop the mmap_lock. Therefore, the value of *locked will still be zero when
1474 * the function returns, even though it may have changed temporarily during
1475 * function execution.
1476 *
1477 * Please note that this function, unlike __get_user_pages(), will not return 0
1478 * for nr_pages > 0, unless FOLL_NOWAIT is used.
1479 */
__get_user_pages_locked(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,struct page ** pages,int * locked,unsigned int flags)1480 static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
1481 unsigned long start,
1482 unsigned long nr_pages,
1483 struct page **pages,
1484 int *locked,
1485 unsigned int flags)
1486 {
1487 long ret, pages_done;
1488 bool must_unlock = false;
1489
1490 /*
1491 * The internal caller expects GUP to manage the lock internally and the
1492 * lock must be released when this returns.
1493 */
1494 if (!*locked) {
1495 if (mmap_read_lock_killable(mm))
1496 return -EAGAIN;
1497 must_unlock = true;
1498 *locked = 1;
1499 }
1500 else
1501 mmap_assert_locked(mm);
1502
1503 if (flags & FOLL_PIN)
1504 mm_set_has_pinned_flag(&mm->flags);
1505
1506 /*
1507 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1508 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1509 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1510 * for FOLL_GET, not for the newer FOLL_PIN.
1511 *
1512 * FOLL_PIN always expects pages to be non-null, but no need to assert
1513 * that here, as any failures will be obvious enough.
1514 */
1515 if (pages && !(flags & FOLL_PIN))
1516 flags |= FOLL_GET;
1517
1518 pages_done = 0;
1519 for (;;) {
1520 ret = __get_user_pages(mm, start, nr_pages, flags, pages,
1521 locked);
1522 if (!(flags & FOLL_UNLOCKABLE)) {
1523 /* VM_FAULT_RETRY couldn't trigger, bypass */
1524 pages_done = ret;
1525 break;
1526 }
1527
1528 /* VM_FAULT_RETRY or VM_FAULT_COMPLETED cannot return errors */
1529 if (!*locked) {
1530 BUG_ON(ret < 0);
1531 BUG_ON(ret >= nr_pages);
1532 }
1533
1534 if (ret > 0) {
1535 nr_pages -= ret;
1536 pages_done += ret;
1537 if (!nr_pages)
1538 break;
1539 }
1540 if (*locked) {
1541 /*
1542 * VM_FAULT_RETRY didn't trigger or it was a
1543 * FOLL_NOWAIT.
1544 */
1545 if (!pages_done)
1546 pages_done = ret;
1547 break;
1548 }
1549 /*
1550 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1551 * For the prefault case (!pages) we only update counts.
1552 */
1553 if (likely(pages))
1554 pages += ret;
1555 start += ret << PAGE_SHIFT;
1556
1557 /* The lock was temporarily dropped, so we must unlock later */
1558 must_unlock = true;
1559
1560 retry:
1561 /*
1562 * Repeat on the address that fired VM_FAULT_RETRY
1563 * with both FAULT_FLAG_ALLOW_RETRY and
1564 * FAULT_FLAG_TRIED. Note that GUP can be interrupted
1565 * by fatal signals of even common signals, depending on
1566 * the caller's request. So we need to check it before we
1567 * start trying again otherwise it can loop forever.
1568 */
1569 if (gup_signal_pending(flags)) {
1570 if (!pages_done)
1571 pages_done = -EINTR;
1572 break;
1573 }
1574
1575 ret = mmap_read_lock_killable(mm);
1576 if (ret) {
1577 BUG_ON(ret > 0);
1578 if (!pages_done)
1579 pages_done = ret;
1580 break;
1581 }
1582
1583 *locked = 1;
1584 ret = __get_user_pages(mm, start, 1, flags | FOLL_TRIED,
1585 pages, locked);
1586 if (!*locked) {
1587 /* Continue to retry until we succeeded */
1588 BUG_ON(ret != 0);
1589 goto retry;
1590 }
1591 if (ret != 1) {
1592 BUG_ON(ret > 1);
1593 if (!pages_done)
1594 pages_done = ret;
1595 break;
1596 }
1597 nr_pages--;
1598 pages_done++;
1599 if (!nr_pages)
1600 break;
1601 if (likely(pages))
1602 pages++;
1603 start += PAGE_SIZE;
1604 }
1605 if (must_unlock && *locked) {
1606 /*
1607 * We either temporarily dropped the lock, or the caller
1608 * requested that we both acquire and drop the lock. Either way,
1609 * we must now unlock, and notify the caller of that state.
1610 */
1611 mmap_read_unlock(mm);
1612 *locked = 0;
1613 }
1614 return pages_done;
1615 }
1616
1617 /**
1618 * populate_vma_page_range() - populate a range of pages in the vma.
1619 * @vma: target vma
1620 * @start: start address
1621 * @end: end address
1622 * @locked: whether the mmap_lock is still held
1623 *
1624 * This takes care of mlocking the pages too if VM_LOCKED is set.
1625 *
1626 * Return either number of pages pinned in the vma, or a negative error
1627 * code on error.
1628 *
1629 * vma->vm_mm->mmap_lock must be held.
1630 *
1631 * If @locked is NULL, it may be held for read or write and will
1632 * be unperturbed.
1633 *
1634 * If @locked is non-NULL, it must held for read only and may be
1635 * released. If it's released, *@locked will be set to 0.
1636 */
populate_vma_page_range(struct vm_area_struct * vma,unsigned long start,unsigned long end,int * locked)1637 long populate_vma_page_range(struct vm_area_struct *vma,
1638 unsigned long start, unsigned long end, int *locked)
1639 {
1640 struct mm_struct *mm = vma->vm_mm;
1641 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1642 int local_locked = 1;
1643 int gup_flags;
1644 long ret;
1645
1646 VM_BUG_ON(!PAGE_ALIGNED(start));
1647 VM_BUG_ON(!PAGE_ALIGNED(end));
1648 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1649 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1650 mmap_assert_locked(mm);
1651
1652 /*
1653 * Rightly or wrongly, the VM_LOCKONFAULT case has never used
1654 * faultin_page() to break COW, so it has no work to do here.
1655 */
1656 if (vma->vm_flags & VM_LOCKONFAULT)
1657 return nr_pages;
1658
1659 gup_flags = FOLL_TOUCH;
1660 /*
1661 * We want to touch writable mappings with a write fault in order
1662 * to break COW, except for shared mappings because these don't COW
1663 * and we would not want to dirty them for nothing.
1664 */
1665 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1666 gup_flags |= FOLL_WRITE;
1667
1668 /*
1669 * We want mlock to succeed for regions that have any permissions
1670 * other than PROT_NONE.
1671 */
1672 if (vma_is_accessible(vma))
1673 gup_flags |= FOLL_FORCE;
1674
1675 if (locked)
1676 gup_flags |= FOLL_UNLOCKABLE;
1677
1678 /*
1679 * We made sure addr is within a VMA, so the following will
1680 * not result in a stack expansion that recurses back here.
1681 */
1682 ret = __get_user_pages(mm, start, nr_pages, gup_flags,
1683 NULL, locked ? locked : &local_locked);
1684 lru_add_drain();
1685 return ret;
1686 }
1687
1688 /*
1689 * faultin_page_range() - populate (prefault) page tables inside the
1690 * given range readable/writable
1691 *
1692 * This takes care of mlocking the pages, too, if VM_LOCKED is set.
1693 *
1694 * @mm: the mm to populate page tables in
1695 * @start: start address
1696 * @end: end address
1697 * @write: whether to prefault readable or writable
1698 * @locked: whether the mmap_lock is still held
1699 *
1700 * Returns either number of processed pages in the MM, or a negative error
1701 * code on error (see __get_user_pages()). Note that this function reports
1702 * errors related to VMAs, such as incompatible mappings, as expected by
1703 * MADV_POPULATE_(READ|WRITE).
1704 *
1705 * The range must be page-aligned.
1706 *
1707 * mm->mmap_lock must be held. If it's released, *@locked will be set to 0.
1708 */
faultin_page_range(struct mm_struct * mm,unsigned long start,unsigned long end,bool write,int * locked)1709 long faultin_page_range(struct mm_struct *mm, unsigned long start,
1710 unsigned long end, bool write, int *locked)
1711 {
1712 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1713 int gup_flags;
1714 long ret;
1715
1716 VM_BUG_ON(!PAGE_ALIGNED(start));
1717 VM_BUG_ON(!PAGE_ALIGNED(end));
1718 mmap_assert_locked(mm);
1719
1720 /*
1721 * FOLL_TOUCH: Mark page accessed and thereby young; will also mark
1722 * the page dirty with FOLL_WRITE -- which doesn't make a
1723 * difference with !FOLL_FORCE, because the page is writable
1724 * in the page table.
1725 * FOLL_HWPOISON: Return -EHWPOISON instead of -EFAULT when we hit
1726 * a poisoned page.
1727 * !FOLL_FORCE: Require proper access permissions.
1728 */
1729 gup_flags = FOLL_TOUCH | FOLL_HWPOISON | FOLL_UNLOCKABLE |
1730 FOLL_MADV_POPULATE;
1731 if (write)
1732 gup_flags |= FOLL_WRITE;
1733
1734 ret = __get_user_pages_locked(mm, start, nr_pages, NULL, locked,
1735 gup_flags);
1736 lru_add_drain();
1737 return ret;
1738 }
1739
1740 /*
1741 * __mm_populate - populate and/or mlock pages within a range of address space.
1742 *
1743 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1744 * flags. VMAs must be already marked with the desired vm_flags, and
1745 * mmap_lock must not be held.
1746 */
__mm_populate(unsigned long start,unsigned long len,int ignore_errors)1747 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1748 {
1749 struct mm_struct *mm = current->mm;
1750 unsigned long end, nstart, nend;
1751 struct vm_area_struct *vma = NULL;
1752 int locked = 0;
1753 long ret = 0;
1754
1755 end = start + len;
1756
1757 for (nstart = start; nstart < end; nstart = nend) {
1758 /*
1759 * We want to fault in pages for [nstart; end) address range.
1760 * Find first corresponding VMA.
1761 */
1762 if (!locked) {
1763 locked = 1;
1764 mmap_read_lock(mm);
1765 vma = find_vma_intersection(mm, nstart, end);
1766 } else if (nstart >= vma->vm_end)
1767 vma = find_vma_intersection(mm, vma->vm_end, end);
1768
1769 if (!vma)
1770 break;
1771 /*
1772 * Set [nstart; nend) to intersection of desired address
1773 * range with the first VMA. Also, skip undesirable VMA types.
1774 */
1775 nend = min(end, vma->vm_end);
1776 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1777 continue;
1778 if (nstart < vma->vm_start)
1779 nstart = vma->vm_start;
1780 /*
1781 * Now fault in a range of pages. populate_vma_page_range()
1782 * double checks the vma flags, so that it won't mlock pages
1783 * if the vma was already munlocked.
1784 */
1785 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1786 if (ret < 0) {
1787 if (ignore_errors) {
1788 ret = 0;
1789 continue; /* continue at next VMA */
1790 }
1791 break;
1792 }
1793 nend = nstart + ret * PAGE_SIZE;
1794 ret = 0;
1795 }
1796 if (locked)
1797 mmap_read_unlock(mm);
1798 return ret; /* 0 or negative error code */
1799 }
1800 #else /* CONFIG_MMU */
__get_user_pages_locked(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,struct page ** pages,int * locked,unsigned int foll_flags)1801 static long __get_user_pages_locked(struct mm_struct *mm, unsigned long start,
1802 unsigned long nr_pages, struct page **pages,
1803 int *locked, unsigned int foll_flags)
1804 {
1805 struct vm_area_struct *vma;
1806 bool must_unlock = false;
1807 unsigned long vm_flags;
1808 long i;
1809
1810 if (!nr_pages)
1811 return 0;
1812
1813 /*
1814 * The internal caller expects GUP to manage the lock internally and the
1815 * lock must be released when this returns.
1816 */
1817 if (!*locked) {
1818 if (mmap_read_lock_killable(mm))
1819 return -EAGAIN;
1820 must_unlock = true;
1821 *locked = 1;
1822 }
1823
1824 /* calculate required read or write permissions.
1825 * If FOLL_FORCE is set, we only require the "MAY" flags.
1826 */
1827 vm_flags = (foll_flags & FOLL_WRITE) ?
1828 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1829 vm_flags &= (foll_flags & FOLL_FORCE) ?
1830 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1831
1832 for (i = 0; i < nr_pages; i++) {
1833 vma = find_vma(mm, start);
1834 if (!vma)
1835 break;
1836
1837 /* protect what we can, including chardevs */
1838 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1839 !(vm_flags & vma->vm_flags))
1840 break;
1841
1842 if (pages) {
1843 pages[i] = virt_to_page((void *)start);
1844 if (pages[i])
1845 get_page(pages[i]);
1846 }
1847
1848 start = (start + PAGE_SIZE) & PAGE_MASK;
1849 }
1850
1851 if (must_unlock && *locked) {
1852 mmap_read_unlock(mm);
1853 *locked = 0;
1854 }
1855
1856 return i ? : -EFAULT;
1857 }
1858 #endif /* !CONFIG_MMU */
1859
1860 /**
1861 * fault_in_writeable - fault in userspace address range for writing
1862 * @uaddr: start of address range
1863 * @size: size of address range
1864 *
1865 * Returns the number of bytes not faulted in (like copy_to_user() and
1866 * copy_from_user()).
1867 */
fault_in_writeable(char __user * uaddr,size_t size)1868 size_t fault_in_writeable(char __user *uaddr, size_t size)
1869 {
1870 char __user *start = uaddr, *end;
1871
1872 if (unlikely(size == 0))
1873 return 0;
1874 if (!user_write_access_begin(uaddr, size))
1875 return size;
1876 if (!PAGE_ALIGNED(uaddr)) {
1877 unsafe_put_user(0, uaddr, out);
1878 uaddr = (char __user *)PAGE_ALIGN((unsigned long)uaddr);
1879 }
1880 end = (char __user *)PAGE_ALIGN((unsigned long)start + size);
1881 if (unlikely(end < start))
1882 end = NULL;
1883 while (uaddr != end) {
1884 unsafe_put_user(0, uaddr, out);
1885 uaddr += PAGE_SIZE;
1886 }
1887
1888 out:
1889 user_write_access_end();
1890 if (size > uaddr - start)
1891 return size - (uaddr - start);
1892 return 0;
1893 }
1894 EXPORT_SYMBOL(fault_in_writeable);
1895
1896 /**
1897 * fault_in_subpage_writeable - fault in an address range for writing
1898 * @uaddr: start of address range
1899 * @size: size of address range
1900 *
1901 * Fault in a user address range for writing while checking for permissions at
1902 * sub-page granularity (e.g. arm64 MTE). This function should be used when
1903 * the caller cannot guarantee forward progress of a copy_to_user() loop.
1904 *
1905 * Returns the number of bytes not faulted in (like copy_to_user() and
1906 * copy_from_user()).
1907 */
fault_in_subpage_writeable(char __user * uaddr,size_t size)1908 size_t fault_in_subpage_writeable(char __user *uaddr, size_t size)
1909 {
1910 size_t faulted_in;
1911
1912 /*
1913 * Attempt faulting in at page granularity first for page table
1914 * permission checking. The arch-specific probe_subpage_writeable()
1915 * functions may not check for this.
1916 */
1917 faulted_in = size - fault_in_writeable(uaddr, size);
1918 if (faulted_in)
1919 faulted_in -= probe_subpage_writeable(uaddr, faulted_in);
1920
1921 return size - faulted_in;
1922 }
1923 EXPORT_SYMBOL(fault_in_subpage_writeable);
1924
1925 /*
1926 * fault_in_safe_writeable - fault in an address range for writing
1927 * @uaddr: start of address range
1928 * @size: length of address range
1929 *
1930 * Faults in an address range for writing. This is primarily useful when we
1931 * already know that some or all of the pages in the address range aren't in
1932 * memory.
1933 *
1934 * Unlike fault_in_writeable(), this function is non-destructive.
1935 *
1936 * Note that we don't pin or otherwise hold the pages referenced that we fault
1937 * in. There's no guarantee that they'll stay in memory for any duration of
1938 * time.
1939 *
1940 * Returns the number of bytes not faulted in, like copy_to_user() and
1941 * copy_from_user().
1942 */
fault_in_safe_writeable(const char __user * uaddr,size_t size)1943 size_t fault_in_safe_writeable(const char __user *uaddr, size_t size)
1944 {
1945 unsigned long start = (unsigned long)uaddr, end;
1946 struct mm_struct *mm = current->mm;
1947 bool unlocked = false;
1948
1949 if (unlikely(size == 0))
1950 return 0;
1951 end = PAGE_ALIGN(start + size);
1952 if (end < start)
1953 end = 0;
1954
1955 mmap_read_lock(mm);
1956 do {
1957 if (fixup_user_fault(mm, start, FAULT_FLAG_WRITE, &unlocked))
1958 break;
1959 start = (start + PAGE_SIZE) & PAGE_MASK;
1960 } while (start != end);
1961 mmap_read_unlock(mm);
1962
1963 if (size > (unsigned long)uaddr - start)
1964 return size - ((unsigned long)uaddr - start);
1965 return 0;
1966 }
1967 EXPORT_SYMBOL(fault_in_safe_writeable);
1968
1969 /**
1970 * fault_in_readable - fault in userspace address range for reading
1971 * @uaddr: start of user address range
1972 * @size: size of user address range
1973 *
1974 * Returns the number of bytes not faulted in (like copy_to_user() and
1975 * copy_from_user()).
1976 */
fault_in_readable(const char __user * uaddr,size_t size)1977 size_t fault_in_readable(const char __user *uaddr, size_t size)
1978 {
1979 const char __user *start = uaddr, *end;
1980 volatile char c;
1981
1982 if (unlikely(size == 0))
1983 return 0;
1984 if (!user_read_access_begin(uaddr, size))
1985 return size;
1986 if (!PAGE_ALIGNED(uaddr)) {
1987 unsafe_get_user(c, uaddr, out);
1988 uaddr = (const char __user *)PAGE_ALIGN((unsigned long)uaddr);
1989 }
1990 end = (const char __user *)PAGE_ALIGN((unsigned long)start + size);
1991 if (unlikely(end < start))
1992 end = NULL;
1993 while (uaddr != end) {
1994 unsafe_get_user(c, uaddr, out);
1995 uaddr += PAGE_SIZE;
1996 }
1997
1998 out:
1999 user_read_access_end();
2000 (void)c;
2001 if (size > uaddr - start)
2002 return size - (uaddr - start);
2003 return 0;
2004 }
2005 EXPORT_SYMBOL(fault_in_readable);
2006
2007 /**
2008 * get_dump_page() - pin user page in memory while writing it to core dump
2009 * @addr: user address
2010 *
2011 * Returns struct page pointer of user page pinned for dump,
2012 * to be freed afterwards by put_page().
2013 *
2014 * Returns NULL on any kind of failure - a hole must then be inserted into
2015 * the corefile, to preserve alignment with its headers; and also returns
2016 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2017 * allowing a hole to be left in the corefile to save disk space.
2018 *
2019 * Called without mmap_lock (takes and releases the mmap_lock by itself).
2020 */
2021 #ifdef CONFIG_ELF_CORE
get_dump_page(unsigned long addr)2022 struct page *get_dump_page(unsigned long addr)
2023 {
2024 struct page *page;
2025 int locked = 0;
2026 int ret;
2027
2028 ret = __get_user_pages_locked(current->mm, addr, 1, &page, &locked,
2029 FOLL_FORCE | FOLL_DUMP | FOLL_GET);
2030 return (ret == 1) ? page : NULL;
2031 }
2032 #endif /* CONFIG_ELF_CORE */
2033
2034 #ifdef CONFIG_MIGRATION
2035 /*
2036 * Returns the number of collected pages. Return value is always >= 0.
2037 */
collect_longterm_unpinnable_pages(struct list_head * movable_page_list,unsigned long nr_pages,struct page ** pages)2038 static unsigned long collect_longterm_unpinnable_pages(
2039 struct list_head *movable_page_list,
2040 unsigned long nr_pages,
2041 struct page **pages)
2042 {
2043 unsigned long i, collected = 0;
2044 struct folio *prev_folio = NULL;
2045 bool drain_allow = true;
2046
2047 for (i = 0; i < nr_pages; i++) {
2048 struct folio *folio = page_folio(pages[i]);
2049
2050 if (folio == prev_folio)
2051 continue;
2052 prev_folio = folio;
2053
2054 if (folio_is_longterm_pinnable(folio))
2055 continue;
2056
2057 collected++;
2058
2059 if (folio_is_device_coherent(folio))
2060 continue;
2061
2062 if (folio_test_hugetlb(folio)) {
2063 isolate_hugetlb(folio, movable_page_list);
2064 continue;
2065 }
2066
2067 if (!folio_test_lru(folio) && drain_allow) {
2068 lru_add_drain_all();
2069 drain_allow = false;
2070 }
2071
2072 if (!folio_isolate_lru(folio))
2073 continue;
2074
2075 list_add_tail(&folio->lru, movable_page_list);
2076 node_stat_mod_folio(folio,
2077 NR_ISOLATED_ANON + folio_is_file_lru(folio),
2078 folio_nr_pages(folio));
2079 }
2080
2081 return collected;
2082 }
2083
2084 /*
2085 * Unpins all pages and migrates device coherent pages and movable_page_list.
2086 * Returns -EAGAIN if all pages were successfully migrated or -errno for failure
2087 * (or partial success).
2088 */
migrate_longterm_unpinnable_pages(struct list_head * movable_page_list,unsigned long nr_pages,struct page ** pages)2089 static int migrate_longterm_unpinnable_pages(
2090 struct list_head *movable_page_list,
2091 unsigned long nr_pages,
2092 struct page **pages)
2093 {
2094 int ret;
2095 unsigned long i;
2096
2097 for (i = 0; i < nr_pages; i++) {
2098 struct folio *folio = page_folio(pages[i]);
2099
2100 if (folio_is_device_coherent(folio)) {
2101 /*
2102 * Migration will fail if the page is pinned, so convert
2103 * the pin on the source page to a normal reference.
2104 */
2105 pages[i] = NULL;
2106 folio_get(folio);
2107 gup_put_folio(folio, 1, FOLL_PIN);
2108
2109 if (migrate_device_coherent_page(&folio->page)) {
2110 ret = -EBUSY;
2111 goto err;
2112 }
2113
2114 continue;
2115 }
2116
2117 /*
2118 * We can't migrate pages with unexpected references, so drop
2119 * the reference obtained by __get_user_pages_locked().
2120 * Migrating pages have been added to movable_page_list after
2121 * calling folio_isolate_lru() which takes a reference so the
2122 * page won't be freed if it's migrating.
2123 */
2124 unpin_user_page(pages[i]);
2125 pages[i] = NULL;
2126 }
2127
2128 if (!list_empty(movable_page_list)) {
2129 struct migration_target_control mtc = {
2130 .nid = NUMA_NO_NODE,
2131 .gfp_mask = GFP_USER | __GFP_NOWARN,
2132 };
2133
2134 if (migrate_pages(movable_page_list, alloc_migration_target,
2135 NULL, (unsigned long)&mtc, MIGRATE_SYNC,
2136 MR_LONGTERM_PIN, NULL)) {
2137 ret = -ENOMEM;
2138 goto err;
2139 }
2140 }
2141
2142 putback_movable_pages(movable_page_list);
2143
2144 return -EAGAIN;
2145
2146 err:
2147 for (i = 0; i < nr_pages; i++)
2148 if (pages[i])
2149 unpin_user_page(pages[i]);
2150 putback_movable_pages(movable_page_list);
2151
2152 return ret;
2153 }
2154
2155 /*
2156 * Check whether all pages are *allowed* to be pinned. Rather confusingly, all
2157 * pages in the range are required to be pinned via FOLL_PIN, before calling
2158 * this routine.
2159 *
2160 * If any pages in the range are not allowed to be pinned, then this routine
2161 * will migrate those pages away, unpin all the pages in the range and return
2162 * -EAGAIN. The caller should re-pin the entire range with FOLL_PIN and then
2163 * call this routine again.
2164 *
2165 * If an error other than -EAGAIN occurs, this indicates a migration failure.
2166 * The caller should give up, and propagate the error back up the call stack.
2167 *
2168 * If everything is OK and all pages in the range are allowed to be pinned, then
2169 * this routine leaves all pages pinned and returns zero for success.
2170 */
check_and_migrate_movable_pages(unsigned long nr_pages,struct page ** pages)2171 static long check_and_migrate_movable_pages(unsigned long nr_pages,
2172 struct page **pages)
2173 {
2174 unsigned long collected;
2175 LIST_HEAD(movable_page_list);
2176
2177 collected = collect_longterm_unpinnable_pages(&movable_page_list,
2178 nr_pages, pages);
2179 if (!collected)
2180 return 0;
2181
2182 return migrate_longterm_unpinnable_pages(&movable_page_list, nr_pages,
2183 pages);
2184 }
2185 #else
check_and_migrate_movable_pages(unsigned long nr_pages,struct page ** pages)2186 static long check_and_migrate_movable_pages(unsigned long nr_pages,
2187 struct page **pages)
2188 {
2189 return 0;
2190 }
2191 #endif /* CONFIG_MIGRATION */
2192
2193 /*
2194 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
2195 * allows us to process the FOLL_LONGTERM flag.
2196 */
__gup_longterm_locked(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,struct page ** pages,int * locked,unsigned int gup_flags)2197 static long __gup_longterm_locked(struct mm_struct *mm,
2198 unsigned long start,
2199 unsigned long nr_pages,
2200 struct page **pages,
2201 int *locked,
2202 unsigned int gup_flags)
2203 {
2204 unsigned int flags;
2205 long rc, nr_pinned_pages;
2206
2207 if (!(gup_flags & FOLL_LONGTERM))
2208 return __get_user_pages_locked(mm, start, nr_pages, pages,
2209 locked, gup_flags);
2210
2211 flags = memalloc_pin_save();
2212 do {
2213 nr_pinned_pages = __get_user_pages_locked(mm, start, nr_pages,
2214 pages, locked,
2215 gup_flags);
2216 if (nr_pinned_pages <= 0) {
2217 rc = nr_pinned_pages;
2218 break;
2219 }
2220
2221 /* FOLL_LONGTERM implies FOLL_PIN */
2222 rc = check_and_migrate_movable_pages(nr_pinned_pages, pages);
2223 } while (rc == -EAGAIN);
2224 memalloc_pin_restore(flags);
2225 return rc ? rc : nr_pinned_pages;
2226 }
2227
2228 /*
2229 * Check that the given flags are valid for the exported gup/pup interface, and
2230 * update them with the required flags that the caller must have set.
2231 */
is_valid_gup_args(struct page ** pages,int * locked,unsigned int * gup_flags_p,unsigned int to_set)2232 static bool is_valid_gup_args(struct page **pages, int *locked,
2233 unsigned int *gup_flags_p, unsigned int to_set)
2234 {
2235 unsigned int gup_flags = *gup_flags_p;
2236
2237 /*
2238 * These flags not allowed to be specified externally to the gup
2239 * interfaces:
2240 * - FOLL_TOUCH/FOLL_PIN/FOLL_TRIED/FOLL_FAST_ONLY are internal only
2241 * - FOLL_REMOTE is internal only and used on follow_page()
2242 * - FOLL_UNLOCKABLE is internal only and used if locked is !NULL
2243 */
2244 if (WARN_ON_ONCE(gup_flags & INTERNAL_GUP_FLAGS))
2245 return false;
2246
2247 gup_flags |= to_set;
2248 if (locked) {
2249 /* At the external interface locked must be set */
2250 if (WARN_ON_ONCE(*locked != 1))
2251 return false;
2252
2253 gup_flags |= FOLL_UNLOCKABLE;
2254 }
2255
2256 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2257 if (WARN_ON_ONCE((gup_flags & (FOLL_PIN | FOLL_GET)) ==
2258 (FOLL_PIN | FOLL_GET)))
2259 return false;
2260
2261 /* LONGTERM can only be specified when pinning */
2262 if (WARN_ON_ONCE(!(gup_flags & FOLL_PIN) && (gup_flags & FOLL_LONGTERM)))
2263 return false;
2264
2265 /* Pages input must be given if using GET/PIN */
2266 if (WARN_ON_ONCE((gup_flags & (FOLL_GET | FOLL_PIN)) && !pages))
2267 return false;
2268
2269 /* We want to allow the pgmap to be hot-unplugged at all times */
2270 if (WARN_ON_ONCE((gup_flags & FOLL_LONGTERM) &&
2271 (gup_flags & FOLL_PCI_P2PDMA)))
2272 return false;
2273
2274 *gup_flags_p = gup_flags;
2275 return true;
2276 }
2277
2278 #ifdef CONFIG_MMU
2279 /**
2280 * get_user_pages_remote() - pin user pages in memory
2281 * @mm: mm_struct of target mm
2282 * @start: starting user address
2283 * @nr_pages: number of pages from start to pin
2284 * @gup_flags: flags modifying lookup behaviour
2285 * @pages: array that receives pointers to the pages pinned.
2286 * Should be at least nr_pages long. Or NULL, if caller
2287 * only intends to ensure the pages are faulted in.
2288 * @locked: pointer to lock flag indicating whether lock is held and
2289 * subsequently whether VM_FAULT_RETRY functionality can be
2290 * utilised. Lock must initially be held.
2291 *
2292 * Returns either number of pages pinned (which may be less than the
2293 * number requested), or an error. Details about the return value:
2294 *
2295 * -- If nr_pages is 0, returns 0.
2296 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
2297 * -- If nr_pages is >0, and some pages were pinned, returns the number of
2298 * pages pinned. Again, this may be less than nr_pages.
2299 *
2300 * The caller is responsible for releasing returned @pages, via put_page().
2301 *
2302 * Must be called with mmap_lock held for read or write.
2303 *
2304 * get_user_pages_remote walks a process's page tables and takes a reference
2305 * to each struct page that each user address corresponds to at a given
2306 * instant. That is, it takes the page that would be accessed if a user
2307 * thread accesses the given user virtual address at that instant.
2308 *
2309 * This does not guarantee that the page exists in the user mappings when
2310 * get_user_pages_remote returns, and there may even be a completely different
2311 * page there in some cases (eg. if mmapped pagecache has been invalidated
2312 * and subsequently re-faulted). However it does guarantee that the page
2313 * won't be freed completely. And mostly callers simply care that the page
2314 * contains data that was valid *at some point in time*. Typically, an IO
2315 * or similar operation cannot guarantee anything stronger anyway because
2316 * locks can't be held over the syscall boundary.
2317 *
2318 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
2319 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
2320 * be called after the page is finished with, and before put_page is called.
2321 *
2322 * get_user_pages_remote is typically used for fewer-copy IO operations,
2323 * to get a handle on the memory by some means other than accesses
2324 * via the user virtual addresses. The pages may be submitted for
2325 * DMA to devices or accessed via their kernel linear mapping (via the
2326 * kmap APIs). Care should be taken to use the correct cache flushing APIs.
2327 *
2328 * See also get_user_pages_fast, for performance critical applications.
2329 *
2330 * get_user_pages_remote should be phased out in favor of
2331 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
2332 * should use get_user_pages_remote because it cannot pass
2333 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
2334 */
get_user_pages_remote(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,int * locked)2335 long get_user_pages_remote(struct mm_struct *mm,
2336 unsigned long start, unsigned long nr_pages,
2337 unsigned int gup_flags, struct page **pages,
2338 int *locked)
2339 {
2340 int local_locked = 1;
2341
2342 if (!is_valid_gup_args(pages, locked, &gup_flags,
2343 FOLL_TOUCH | FOLL_REMOTE))
2344 return -EINVAL;
2345
2346 return __get_user_pages_locked(mm, start, nr_pages, pages,
2347 locked ? locked : &local_locked,
2348 gup_flags);
2349 }
2350 EXPORT_SYMBOL(get_user_pages_remote);
2351
2352 #else /* CONFIG_MMU */
get_user_pages_remote(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,int * locked)2353 long get_user_pages_remote(struct mm_struct *mm,
2354 unsigned long start, unsigned long nr_pages,
2355 unsigned int gup_flags, struct page **pages,
2356 int *locked)
2357 {
2358 return 0;
2359 }
2360 #endif /* !CONFIG_MMU */
2361
2362 /**
2363 * get_user_pages() - pin user pages in memory
2364 * @start: starting user address
2365 * @nr_pages: number of pages from start to pin
2366 * @gup_flags: flags modifying lookup behaviour
2367 * @pages: array that receives pointers to the pages pinned.
2368 * Should be at least nr_pages long. Or NULL, if caller
2369 * only intends to ensure the pages are faulted in.
2370 *
2371 * This is the same as get_user_pages_remote(), just with a less-flexible
2372 * calling convention where we assume that the mm being operated on belongs to
2373 * the current task, and doesn't allow passing of a locked parameter. We also
2374 * obviously don't pass FOLL_REMOTE in here.
2375 */
get_user_pages(unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages)2376 long get_user_pages(unsigned long start, unsigned long nr_pages,
2377 unsigned int gup_flags, struct page **pages)
2378 {
2379 int locked = 1;
2380
2381 if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_TOUCH))
2382 return -EINVAL;
2383
2384 return __get_user_pages_locked(current->mm, start, nr_pages, pages,
2385 &locked, gup_flags);
2386 }
2387 EXPORT_SYMBOL(get_user_pages);
2388
2389 /*
2390 * get_user_pages_unlocked() is suitable to replace the form:
2391 *
2392 * mmap_read_lock(mm);
2393 * get_user_pages(mm, ..., pages, NULL);
2394 * mmap_read_unlock(mm);
2395 *
2396 * with:
2397 *
2398 * get_user_pages_unlocked(mm, ..., pages);
2399 *
2400 * It is functionally equivalent to get_user_pages_fast so
2401 * get_user_pages_fast should be used instead if specific gup_flags
2402 * (e.g. FOLL_FORCE) are not required.
2403 */
get_user_pages_unlocked(unsigned long start,unsigned long nr_pages,struct page ** pages,unsigned int gup_flags)2404 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
2405 struct page **pages, unsigned int gup_flags)
2406 {
2407 int locked = 0;
2408
2409 if (!is_valid_gup_args(pages, NULL, &gup_flags,
2410 FOLL_TOUCH | FOLL_UNLOCKABLE))
2411 return -EINVAL;
2412
2413 return __get_user_pages_locked(current->mm, start, nr_pages, pages,
2414 &locked, gup_flags);
2415 }
2416 EXPORT_SYMBOL(get_user_pages_unlocked);
2417
2418 /*
2419 * Fast GUP
2420 *
2421 * get_user_pages_fast attempts to pin user pages by walking the page
2422 * tables directly and avoids taking locks. Thus the walker needs to be
2423 * protected from page table pages being freed from under it, and should
2424 * block any THP splits.
2425 *
2426 * One way to achieve this is to have the walker disable interrupts, and
2427 * rely on IPIs from the TLB flushing code blocking before the page table
2428 * pages are freed. This is unsuitable for architectures that do not need
2429 * to broadcast an IPI when invalidating TLBs.
2430 *
2431 * Another way to achieve this is to batch up page table containing pages
2432 * belonging to more than one mm_user, then rcu_sched a callback to free those
2433 * pages. Disabling interrupts will allow the fast_gup walker to both block
2434 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2435 * (which is a relatively rare event). The code below adopts this strategy.
2436 *
2437 * Before activating this code, please be aware that the following assumptions
2438 * are currently made:
2439 *
2440 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2441 * free pages containing page tables or TLB flushing requires IPI broadcast.
2442 *
2443 * *) ptes can be read atomically by the architecture.
2444 *
2445 * *) access_ok is sufficient to validate userspace address ranges.
2446 *
2447 * The last two assumptions can be relaxed by the addition of helper functions.
2448 *
2449 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2450 */
2451 #ifdef CONFIG_HAVE_FAST_GUP
2452
2453 /*
2454 * Used in the GUP-fast path to determine whether a pin is permitted for a
2455 * specific folio.
2456 *
2457 * This call assumes the caller has pinned the folio, that the lowest page table
2458 * level still points to this folio, and that interrupts have been disabled.
2459 *
2460 * Writing to pinned file-backed dirty tracked folios is inherently problematic
2461 * (see comment describing the writable_file_mapping_allowed() function). We
2462 * therefore try to avoid the most egregious case of a long-term mapping doing
2463 * so.
2464 *
2465 * This function cannot be as thorough as that one as the VMA is not available
2466 * in the fast path, so instead we whitelist known good cases and if in doubt,
2467 * fall back to the slow path.
2468 */
folio_fast_pin_allowed(struct folio * folio,unsigned int flags)2469 static bool folio_fast_pin_allowed(struct folio *folio, unsigned int flags)
2470 {
2471 struct address_space *mapping;
2472 unsigned long mapping_flags;
2473
2474 /*
2475 * If we aren't pinning then no problematic write can occur. A long term
2476 * pin is the most egregious case so this is the one we disallow.
2477 */
2478 if ((flags & (FOLL_PIN | FOLL_LONGTERM | FOLL_WRITE)) !=
2479 (FOLL_PIN | FOLL_LONGTERM | FOLL_WRITE))
2480 return true;
2481
2482 /* The folio is pinned, so we can safely access folio fields. */
2483
2484 if (WARN_ON_ONCE(folio_test_slab(folio)))
2485 return false;
2486
2487 /* hugetlb mappings do not require dirty-tracking. */
2488 if (folio_test_hugetlb(folio))
2489 return true;
2490
2491 /*
2492 * GUP-fast disables IRQs. When IRQS are disabled, RCU grace periods
2493 * cannot proceed, which means no actions performed under RCU can
2494 * proceed either.
2495 *
2496 * inodes and thus their mappings are freed under RCU, which means the
2497 * mapping cannot be freed beneath us and thus we can safely dereference
2498 * it.
2499 */
2500 lockdep_assert_irqs_disabled();
2501
2502 /*
2503 * However, there may be operations which _alter_ the mapping, so ensure
2504 * we read it once and only once.
2505 */
2506 mapping = READ_ONCE(folio->mapping);
2507
2508 /*
2509 * The mapping may have been truncated, in any case we cannot determine
2510 * if this mapping is safe - fall back to slow path to determine how to
2511 * proceed.
2512 */
2513 if (!mapping)
2514 return false;
2515
2516 /* Anonymous folios pose no problem. */
2517 mapping_flags = (unsigned long)mapping & PAGE_MAPPING_FLAGS;
2518 if (mapping_flags)
2519 return mapping_flags & PAGE_MAPPING_ANON;
2520
2521 /*
2522 * At this point, we know the mapping is non-null and points to an
2523 * address_space object. The only remaining whitelisted file system is
2524 * shmem.
2525 */
2526 return shmem_mapping(mapping);
2527 }
2528
undo_dev_pagemap(int * nr,int nr_start,unsigned int flags,struct page ** pages)2529 static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start,
2530 unsigned int flags,
2531 struct page **pages)
2532 {
2533 while ((*nr) - nr_start) {
2534 struct page *page = pages[--(*nr)];
2535
2536 ClearPageReferenced(page);
2537 if (flags & FOLL_PIN)
2538 unpin_user_page(page);
2539 else
2540 put_page(page);
2541 }
2542 }
2543
2544 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2545 /*
2546 * Fast-gup relies on pte change detection to avoid concurrent pgtable
2547 * operations.
2548 *
2549 * To pin the page, fast-gup needs to do below in order:
2550 * (1) pin the page (by prefetching pte), then (2) check pte not changed.
2551 *
2552 * For the rest of pgtable operations where pgtable updates can be racy
2553 * with fast-gup, we need to do (1) clear pte, then (2) check whether page
2554 * is pinned.
2555 *
2556 * Above will work for all pte-level operations, including THP split.
2557 *
2558 * For THP collapse, it's a bit more complicated because fast-gup may be
2559 * walking a pgtable page that is being freed (pte is still valid but pmd
2560 * can be cleared already). To avoid race in such condition, we need to
2561 * also check pmd here to make sure pmd doesn't change (corresponds to
2562 * pmdp_collapse_flush() in the THP collapse code path).
2563 */
gup_pte_range(pmd_t pmd,pmd_t * pmdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2564 static int gup_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr,
2565 unsigned long end, unsigned int flags,
2566 struct page **pages, int *nr)
2567 {
2568 struct dev_pagemap *pgmap = NULL;
2569 int nr_start = *nr, ret = 0;
2570 pte_t *ptep, *ptem;
2571
2572 ptem = ptep = pte_offset_map(&pmd, addr);
2573 if (!ptep)
2574 return 0;
2575 do {
2576 pte_t pte = ptep_get_lockless(ptep);
2577 struct page *page;
2578 struct folio *folio;
2579
2580 /*
2581 * Always fallback to ordinary GUP on PROT_NONE-mapped pages:
2582 * pte_access_permitted() better should reject these pages
2583 * either way: otherwise, GUP-fast might succeed in
2584 * cases where ordinary GUP would fail due to VMA access
2585 * permissions.
2586 */
2587 if (pte_protnone(pte))
2588 goto pte_unmap;
2589
2590 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2591 goto pte_unmap;
2592
2593 if (pte_devmap(pte)) {
2594 if (unlikely(flags & FOLL_LONGTERM))
2595 goto pte_unmap;
2596
2597 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
2598 if (unlikely(!pgmap)) {
2599 undo_dev_pagemap(nr, nr_start, flags, pages);
2600 goto pte_unmap;
2601 }
2602 } else if (pte_special(pte))
2603 goto pte_unmap;
2604
2605 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2606 page = pte_page(pte);
2607
2608 folio = try_grab_folio(page, 1, flags);
2609 if (!folio)
2610 goto pte_unmap;
2611
2612 if (unlikely(folio_is_secretmem(folio))) {
2613 gup_put_folio(folio, 1, flags);
2614 goto pte_unmap;
2615 }
2616
2617 if (unlikely(pmd_val(pmd) != pmd_val(*pmdp)) ||
2618 unlikely(pte_val(pte) != pte_val(ptep_get(ptep)))) {
2619 gup_put_folio(folio, 1, flags);
2620 goto pte_unmap;
2621 }
2622
2623 if (!folio_fast_pin_allowed(folio, flags)) {
2624 gup_put_folio(folio, 1, flags);
2625 goto pte_unmap;
2626 }
2627
2628 if (!pte_write(pte) && gup_must_unshare(NULL, flags, page)) {
2629 gup_put_folio(folio, 1, flags);
2630 goto pte_unmap;
2631 }
2632
2633 /*
2634 * We need to make the page accessible if and only if we are
2635 * going to access its content (the FOLL_PIN case). Please
2636 * see Documentation/core-api/pin_user_pages.rst for
2637 * details.
2638 */
2639 if (flags & FOLL_PIN) {
2640 ret = arch_make_page_accessible(page);
2641 if (ret) {
2642 gup_put_folio(folio, 1, flags);
2643 goto pte_unmap;
2644 }
2645 }
2646 folio_set_referenced(folio);
2647 pages[*nr] = page;
2648 (*nr)++;
2649 } while (ptep++, addr += PAGE_SIZE, addr != end);
2650
2651 ret = 1;
2652
2653 pte_unmap:
2654 if (pgmap)
2655 put_dev_pagemap(pgmap);
2656 pte_unmap(ptem);
2657 return ret;
2658 }
2659 #else
2660
2661 /*
2662 * If we can't determine whether or not a pte is special, then fail immediately
2663 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2664 * to be special.
2665 *
2666 * For a futex to be placed on a THP tail page, get_futex_key requires a
2667 * get_user_pages_fast_only implementation that can pin pages. Thus it's still
2668 * useful to have gup_huge_pmd even if we can't operate on ptes.
2669 */
gup_pte_range(pmd_t pmd,pmd_t * pmdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2670 static int gup_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr,
2671 unsigned long end, unsigned int flags,
2672 struct page **pages, int *nr)
2673 {
2674 return 0;
2675 }
2676 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2677
2678 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
__gup_device_huge(unsigned long pfn,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2679 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
2680 unsigned long end, unsigned int flags,
2681 struct page **pages, int *nr)
2682 {
2683 int nr_start = *nr;
2684 struct dev_pagemap *pgmap = NULL;
2685
2686 do {
2687 struct page *page = pfn_to_page(pfn);
2688
2689 pgmap = get_dev_pagemap(pfn, pgmap);
2690 if (unlikely(!pgmap)) {
2691 undo_dev_pagemap(nr, nr_start, flags, pages);
2692 break;
2693 }
2694
2695 if (!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)) {
2696 undo_dev_pagemap(nr, nr_start, flags, pages);
2697 break;
2698 }
2699
2700 SetPageReferenced(page);
2701 pages[*nr] = page;
2702 if (unlikely(try_grab_page(page, flags))) {
2703 undo_dev_pagemap(nr, nr_start, flags, pages);
2704 break;
2705 }
2706 (*nr)++;
2707 pfn++;
2708 } while (addr += PAGE_SIZE, addr != end);
2709
2710 put_dev_pagemap(pgmap);
2711 return addr == end;
2712 }
2713
__gup_device_huge_pmd(pmd_t orig,pmd_t * pmdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2714 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2715 unsigned long end, unsigned int flags,
2716 struct page **pages, int *nr)
2717 {
2718 unsigned long fault_pfn;
2719 int nr_start = *nr;
2720
2721 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2722 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2723 return 0;
2724
2725 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2726 undo_dev_pagemap(nr, nr_start, flags, pages);
2727 return 0;
2728 }
2729 return 1;
2730 }
2731
__gup_device_huge_pud(pud_t orig,pud_t * pudp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2732 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2733 unsigned long end, unsigned int flags,
2734 struct page **pages, int *nr)
2735 {
2736 unsigned long fault_pfn;
2737 int nr_start = *nr;
2738
2739 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2740 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2741 return 0;
2742
2743 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2744 undo_dev_pagemap(nr, nr_start, flags, pages);
2745 return 0;
2746 }
2747 return 1;
2748 }
2749 #else
__gup_device_huge_pmd(pmd_t orig,pmd_t * pmdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2750 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2751 unsigned long end, unsigned int flags,
2752 struct page **pages, int *nr)
2753 {
2754 BUILD_BUG();
2755 return 0;
2756 }
2757
__gup_device_huge_pud(pud_t pud,pud_t * pudp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2758 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
2759 unsigned long end, unsigned int flags,
2760 struct page **pages, int *nr)
2761 {
2762 BUILD_BUG();
2763 return 0;
2764 }
2765 #endif
2766
record_subpages(struct page * page,unsigned long addr,unsigned long end,struct page ** pages)2767 static int record_subpages(struct page *page, unsigned long addr,
2768 unsigned long end, struct page **pages)
2769 {
2770 int nr;
2771
2772 for (nr = 0; addr != end; nr++, addr += PAGE_SIZE)
2773 pages[nr] = nth_page(page, nr);
2774
2775 return nr;
2776 }
2777
2778 #ifdef CONFIG_ARCH_HAS_HUGEPD
hugepte_addr_end(unsigned long addr,unsigned long end,unsigned long sz)2779 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
2780 unsigned long sz)
2781 {
2782 unsigned long __boundary = (addr + sz) & ~(sz-1);
2783 return (__boundary - 1 < end - 1) ? __boundary : end;
2784 }
2785
gup_hugepte(pte_t * ptep,unsigned long sz,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2786 static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
2787 unsigned long end, unsigned int flags,
2788 struct page **pages, int *nr)
2789 {
2790 unsigned long pte_end;
2791 struct page *page;
2792 struct folio *folio;
2793 pte_t pte;
2794 int refs;
2795
2796 pte_end = (addr + sz) & ~(sz-1);
2797 if (pte_end < end)
2798 end = pte_end;
2799
2800 pte = huge_ptep_get(ptep);
2801
2802 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2803 return 0;
2804
2805 /* hugepages are never "special" */
2806 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2807
2808 page = nth_page(pte_page(pte), (addr & (sz - 1)) >> PAGE_SHIFT);
2809 refs = record_subpages(page, addr, end, pages + *nr);
2810
2811 folio = try_grab_folio(page, refs, flags);
2812 if (!folio)
2813 return 0;
2814
2815 if (unlikely(pte_val(pte) != pte_val(ptep_get(ptep)))) {
2816 gup_put_folio(folio, refs, flags);
2817 return 0;
2818 }
2819
2820 if (!folio_fast_pin_allowed(folio, flags)) {
2821 gup_put_folio(folio, refs, flags);
2822 return 0;
2823 }
2824
2825 if (!pte_write(pte) && gup_must_unshare(NULL, flags, &folio->page)) {
2826 gup_put_folio(folio, refs, flags);
2827 return 0;
2828 }
2829
2830 *nr += refs;
2831 folio_set_referenced(folio);
2832 return 1;
2833 }
2834
gup_huge_pd(hugepd_t hugepd,unsigned long addr,unsigned int pdshift,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2835 static int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2836 unsigned int pdshift, unsigned long end, unsigned int flags,
2837 struct page **pages, int *nr)
2838 {
2839 pte_t *ptep;
2840 unsigned long sz = 1UL << hugepd_shift(hugepd);
2841 unsigned long next;
2842
2843 ptep = hugepte_offset(hugepd, addr, pdshift);
2844 do {
2845 next = hugepte_addr_end(addr, end, sz);
2846 if (!gup_hugepte(ptep, sz, addr, end, flags, pages, nr))
2847 return 0;
2848 } while (ptep++, addr = next, addr != end);
2849
2850 return 1;
2851 }
2852 #else
gup_huge_pd(hugepd_t hugepd,unsigned long addr,unsigned int pdshift,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2853 static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2854 unsigned int pdshift, unsigned long end, unsigned int flags,
2855 struct page **pages, int *nr)
2856 {
2857 return 0;
2858 }
2859 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2860
gup_huge_pmd(pmd_t orig,pmd_t * pmdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2861 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2862 unsigned long end, unsigned int flags,
2863 struct page **pages, int *nr)
2864 {
2865 struct page *page;
2866 struct folio *folio;
2867 int refs;
2868
2869 if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
2870 return 0;
2871
2872 if (pmd_devmap(orig)) {
2873 if (unlikely(flags & FOLL_LONGTERM))
2874 return 0;
2875 return __gup_device_huge_pmd(orig, pmdp, addr, end, flags,
2876 pages, nr);
2877 }
2878
2879 page = nth_page(pmd_page(orig), (addr & ~PMD_MASK) >> PAGE_SHIFT);
2880 refs = record_subpages(page, addr, end, pages + *nr);
2881
2882 folio = try_grab_folio(page, refs, flags);
2883 if (!folio)
2884 return 0;
2885
2886 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2887 gup_put_folio(folio, refs, flags);
2888 return 0;
2889 }
2890
2891 if (!folio_fast_pin_allowed(folio, flags)) {
2892 gup_put_folio(folio, refs, flags);
2893 return 0;
2894 }
2895 if (!pmd_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
2896 gup_put_folio(folio, refs, flags);
2897 return 0;
2898 }
2899
2900 *nr += refs;
2901 folio_set_referenced(folio);
2902 return 1;
2903 }
2904
gup_huge_pud(pud_t orig,pud_t * pudp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2905 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2906 unsigned long end, unsigned int flags,
2907 struct page **pages, int *nr)
2908 {
2909 struct page *page;
2910 struct folio *folio;
2911 int refs;
2912
2913 if (!pud_access_permitted(orig, flags & FOLL_WRITE))
2914 return 0;
2915
2916 if (pud_devmap(orig)) {
2917 if (unlikely(flags & FOLL_LONGTERM))
2918 return 0;
2919 return __gup_device_huge_pud(orig, pudp, addr, end, flags,
2920 pages, nr);
2921 }
2922
2923 page = nth_page(pud_page(orig), (addr & ~PUD_MASK) >> PAGE_SHIFT);
2924 refs = record_subpages(page, addr, end, pages + *nr);
2925
2926 folio = try_grab_folio(page, refs, flags);
2927 if (!folio)
2928 return 0;
2929
2930 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2931 gup_put_folio(folio, refs, flags);
2932 return 0;
2933 }
2934
2935 if (!folio_fast_pin_allowed(folio, flags)) {
2936 gup_put_folio(folio, refs, flags);
2937 return 0;
2938 }
2939
2940 if (!pud_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
2941 gup_put_folio(folio, refs, flags);
2942 return 0;
2943 }
2944
2945 *nr += refs;
2946 folio_set_referenced(folio);
2947 return 1;
2948 }
2949
gup_huge_pgd(pgd_t orig,pgd_t * pgdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2950 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
2951 unsigned long end, unsigned int flags,
2952 struct page **pages, int *nr)
2953 {
2954 int refs;
2955 struct page *page;
2956 struct folio *folio;
2957
2958 if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
2959 return 0;
2960
2961 BUILD_BUG_ON(pgd_devmap(orig));
2962
2963 page = nth_page(pgd_page(orig), (addr & ~PGDIR_MASK) >> PAGE_SHIFT);
2964 refs = record_subpages(page, addr, end, pages + *nr);
2965
2966 folio = try_grab_folio(page, refs, flags);
2967 if (!folio)
2968 return 0;
2969
2970 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
2971 gup_put_folio(folio, refs, flags);
2972 return 0;
2973 }
2974
2975 if (!pgd_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
2976 gup_put_folio(folio, refs, flags);
2977 return 0;
2978 }
2979
2980 if (!folio_fast_pin_allowed(folio, flags)) {
2981 gup_put_folio(folio, refs, flags);
2982 return 0;
2983 }
2984
2985 *nr += refs;
2986 folio_set_referenced(folio);
2987 return 1;
2988 }
2989
gup_pmd_range(pud_t * pudp,pud_t pud,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2990 static int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr, unsigned long end,
2991 unsigned int flags, struct page **pages, int *nr)
2992 {
2993 unsigned long next;
2994 pmd_t *pmdp;
2995
2996 pmdp = pmd_offset_lockless(pudp, pud, addr);
2997 do {
2998 pmd_t pmd = pmdp_get_lockless(pmdp);
2999
3000 next = pmd_addr_end(addr, end);
3001 if (!pmd_present(pmd))
3002 return 0;
3003
3004 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
3005 pmd_devmap(pmd))) {
3006 /* See gup_pte_range() */
3007 if (pmd_protnone(pmd))
3008 return 0;
3009
3010 if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
3011 pages, nr))
3012 return 0;
3013
3014 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
3015 /*
3016 * architecture have different format for hugetlbfs
3017 * pmd format and THP pmd format
3018 */
3019 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
3020 PMD_SHIFT, next, flags, pages, nr))
3021 return 0;
3022 } else if (!gup_pte_range(pmd, pmdp, addr, next, flags, pages, nr))
3023 return 0;
3024 } while (pmdp++, addr = next, addr != end);
3025
3026 return 1;
3027 }
3028
gup_pud_range(p4d_t * p4dp,p4d_t p4d,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3029 static int gup_pud_range(p4d_t *p4dp, p4d_t p4d, unsigned long addr, unsigned long end,
3030 unsigned int flags, struct page **pages, int *nr)
3031 {
3032 unsigned long next;
3033 pud_t *pudp;
3034
3035 pudp = pud_offset_lockless(p4dp, p4d, addr);
3036 do {
3037 pud_t pud = READ_ONCE(*pudp);
3038
3039 next = pud_addr_end(addr, end);
3040 if (unlikely(!pud_present(pud)))
3041 return 0;
3042 if (unlikely(pud_huge(pud) || pud_devmap(pud))) {
3043 if (!gup_huge_pud(pud, pudp, addr, next, flags,
3044 pages, nr))
3045 return 0;
3046 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
3047 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
3048 PUD_SHIFT, next, flags, pages, nr))
3049 return 0;
3050 } else if (!gup_pmd_range(pudp, pud, addr, next, flags, pages, nr))
3051 return 0;
3052 } while (pudp++, addr = next, addr != end);
3053
3054 return 1;
3055 }
3056
gup_p4d_range(pgd_t * pgdp,pgd_t pgd,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3057 static int gup_p4d_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr, unsigned long end,
3058 unsigned int flags, struct page **pages, int *nr)
3059 {
3060 unsigned long next;
3061 p4d_t *p4dp;
3062
3063 p4dp = p4d_offset_lockless(pgdp, pgd, addr);
3064 do {
3065 p4d_t p4d = READ_ONCE(*p4dp);
3066
3067 next = p4d_addr_end(addr, end);
3068 if (p4d_none(p4d))
3069 return 0;
3070 BUILD_BUG_ON(p4d_huge(p4d));
3071 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
3072 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
3073 P4D_SHIFT, next, flags, pages, nr))
3074 return 0;
3075 } else if (!gup_pud_range(p4dp, p4d, addr, next, flags, pages, nr))
3076 return 0;
3077 } while (p4dp++, addr = next, addr != end);
3078
3079 return 1;
3080 }
3081
gup_pgd_range(unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3082 static void gup_pgd_range(unsigned long addr, unsigned long end,
3083 unsigned int flags, struct page **pages, int *nr)
3084 {
3085 unsigned long next;
3086 pgd_t *pgdp;
3087
3088 pgdp = pgd_offset(current->mm, addr);
3089 do {
3090 pgd_t pgd = READ_ONCE(*pgdp);
3091
3092 next = pgd_addr_end(addr, end);
3093 if (pgd_none(pgd))
3094 return;
3095 if (unlikely(pgd_huge(pgd))) {
3096 if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
3097 pages, nr))
3098 return;
3099 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
3100 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
3101 PGDIR_SHIFT, next, flags, pages, nr))
3102 return;
3103 } else if (!gup_p4d_range(pgdp, pgd, addr, next, flags, pages, nr))
3104 return;
3105 } while (pgdp++, addr = next, addr != end);
3106 }
3107 #else
gup_pgd_range(unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3108 static inline void gup_pgd_range(unsigned long addr, unsigned long end,
3109 unsigned int flags, struct page **pages, int *nr)
3110 {
3111 }
3112 #endif /* CONFIG_HAVE_FAST_GUP */
3113
3114 #ifndef gup_fast_permitted
3115 /*
3116 * Check if it's allowed to use get_user_pages_fast_only() for the range, or
3117 * we need to fall back to the slow version:
3118 */
gup_fast_permitted(unsigned long start,unsigned long end)3119 static bool gup_fast_permitted(unsigned long start, unsigned long end)
3120 {
3121 return true;
3122 }
3123 #endif
3124
lockless_pages_from_mm(unsigned long start,unsigned long end,unsigned int gup_flags,struct page ** pages)3125 static unsigned long lockless_pages_from_mm(unsigned long start,
3126 unsigned long end,
3127 unsigned int gup_flags,
3128 struct page **pages)
3129 {
3130 unsigned long flags;
3131 int nr_pinned = 0;
3132 unsigned seq;
3133
3134 if (!IS_ENABLED(CONFIG_HAVE_FAST_GUP) ||
3135 !gup_fast_permitted(start, end))
3136 return 0;
3137
3138 if (gup_flags & FOLL_PIN) {
3139 seq = raw_read_seqcount(¤t->mm->write_protect_seq);
3140 if (seq & 1)
3141 return 0;
3142 }
3143
3144 /*
3145 * Disable interrupts. The nested form is used, in order to allow full,
3146 * general purpose use of this routine.
3147 *
3148 * With interrupts disabled, we block page table pages from being freed
3149 * from under us. See struct mmu_table_batch comments in
3150 * include/asm-generic/tlb.h for more details.
3151 *
3152 * We do not adopt an rcu_read_lock() here as we also want to block IPIs
3153 * that come from THPs splitting.
3154 */
3155 local_irq_save(flags);
3156 gup_pgd_range(start, end, gup_flags, pages, &nr_pinned);
3157 local_irq_restore(flags);
3158
3159 /*
3160 * When pinning pages for DMA there could be a concurrent write protect
3161 * from fork() via copy_page_range(), in this case always fail fast GUP.
3162 */
3163 if (gup_flags & FOLL_PIN) {
3164 if (read_seqcount_retry(¤t->mm->write_protect_seq, seq)) {
3165 unpin_user_pages_lockless(pages, nr_pinned);
3166 return 0;
3167 } else {
3168 sanity_check_pinned_pages(pages, nr_pinned);
3169 }
3170 }
3171 return nr_pinned;
3172 }
3173
internal_get_user_pages_fast(unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages)3174 static int internal_get_user_pages_fast(unsigned long start,
3175 unsigned long nr_pages,
3176 unsigned int gup_flags,
3177 struct page **pages)
3178 {
3179 unsigned long len, end;
3180 unsigned long nr_pinned;
3181 int locked = 0;
3182 int ret;
3183
3184 if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM |
3185 FOLL_FORCE | FOLL_PIN | FOLL_GET |
3186 FOLL_FAST_ONLY | FOLL_NOFAULT |
3187 FOLL_PCI_P2PDMA | FOLL_HONOR_NUMA_FAULT)))
3188 return -EINVAL;
3189
3190 if (gup_flags & FOLL_PIN)
3191 mm_set_has_pinned_flag(¤t->mm->flags);
3192
3193 if (!(gup_flags & FOLL_FAST_ONLY))
3194 might_lock_read(¤t->mm->mmap_lock);
3195
3196 start = untagged_addr(start) & PAGE_MASK;
3197 len = nr_pages << PAGE_SHIFT;
3198 if (check_add_overflow(start, len, &end))
3199 return -EOVERFLOW;
3200 if (end > TASK_SIZE_MAX)
3201 return -EFAULT;
3202 if (unlikely(!access_ok((void __user *)start, len)))
3203 return -EFAULT;
3204
3205 nr_pinned = lockless_pages_from_mm(start, end, gup_flags, pages);
3206 if (nr_pinned == nr_pages || gup_flags & FOLL_FAST_ONLY)
3207 return nr_pinned;
3208
3209 /* Slow path: try to get the remaining pages with get_user_pages */
3210 start += nr_pinned << PAGE_SHIFT;
3211 pages += nr_pinned;
3212 ret = __gup_longterm_locked(current->mm, start, nr_pages - nr_pinned,
3213 pages, &locked,
3214 gup_flags | FOLL_TOUCH | FOLL_UNLOCKABLE);
3215 if (ret < 0) {
3216 /*
3217 * The caller has to unpin the pages we already pinned so
3218 * returning -errno is not an option
3219 */
3220 if (nr_pinned)
3221 return nr_pinned;
3222 return ret;
3223 }
3224 return ret + nr_pinned;
3225 }
3226
3227 /**
3228 * get_user_pages_fast_only() - pin user pages in memory
3229 * @start: starting user address
3230 * @nr_pages: number of pages from start to pin
3231 * @gup_flags: flags modifying pin behaviour
3232 * @pages: array that receives pointers to the pages pinned.
3233 * Should be at least nr_pages long.
3234 *
3235 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
3236 * the regular GUP.
3237 *
3238 * If the architecture does not support this function, simply return with no
3239 * pages pinned.
3240 *
3241 * Careful, careful! COW breaking can go either way, so a non-write
3242 * access can get ambiguous page results. If you call this function without
3243 * 'write' set, you'd better be sure that you're ok with that ambiguity.
3244 */
get_user_pages_fast_only(unsigned long start,int nr_pages,unsigned int gup_flags,struct page ** pages)3245 int get_user_pages_fast_only(unsigned long start, int nr_pages,
3246 unsigned int gup_flags, struct page **pages)
3247 {
3248 /*
3249 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
3250 * because gup fast is always a "pin with a +1 page refcount" request.
3251 *
3252 * FOLL_FAST_ONLY is required in order to match the API description of
3253 * this routine: no fall back to regular ("slow") GUP.
3254 */
3255 if (!is_valid_gup_args(pages, NULL, &gup_flags,
3256 FOLL_GET | FOLL_FAST_ONLY))
3257 return -EINVAL;
3258
3259 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
3260 }
3261 EXPORT_SYMBOL_GPL(get_user_pages_fast_only);
3262
3263 /**
3264 * get_user_pages_fast() - pin user pages in memory
3265 * @start: starting user address
3266 * @nr_pages: number of pages from start to pin
3267 * @gup_flags: flags modifying pin behaviour
3268 * @pages: array that receives pointers to the pages pinned.
3269 * Should be at least nr_pages long.
3270 *
3271 * Attempt to pin user pages in memory without taking mm->mmap_lock.
3272 * If not successful, it will fall back to taking the lock and
3273 * calling get_user_pages().
3274 *
3275 * Returns number of pages pinned. This may be fewer than the number requested.
3276 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
3277 * -errno.
3278 */
get_user_pages_fast(unsigned long start,int nr_pages,unsigned int gup_flags,struct page ** pages)3279 int get_user_pages_fast(unsigned long start, int nr_pages,
3280 unsigned int gup_flags, struct page **pages)
3281 {
3282 /*
3283 * The caller may or may not have explicitly set FOLL_GET; either way is
3284 * OK. However, internally (within mm/gup.c), gup fast variants must set
3285 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
3286 * request.
3287 */
3288 if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_GET))
3289 return -EINVAL;
3290 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
3291 }
3292 EXPORT_SYMBOL_GPL(get_user_pages_fast);
3293
3294 /**
3295 * pin_user_pages_fast() - pin user pages in memory without taking locks
3296 *
3297 * @start: starting user address
3298 * @nr_pages: number of pages from start to pin
3299 * @gup_flags: flags modifying pin behaviour
3300 * @pages: array that receives pointers to the pages pinned.
3301 * Should be at least nr_pages long.
3302 *
3303 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
3304 * get_user_pages_fast() for documentation on the function arguments, because
3305 * the arguments here are identical.
3306 *
3307 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3308 * see Documentation/core-api/pin_user_pages.rst for further details.
3309 *
3310 * Note that if a zero_page is amongst the returned pages, it will not have
3311 * pins in it and unpin_user_page() will not remove pins from it.
3312 */
pin_user_pages_fast(unsigned long start,int nr_pages,unsigned int gup_flags,struct page ** pages)3313 int pin_user_pages_fast(unsigned long start, int nr_pages,
3314 unsigned int gup_flags, struct page **pages)
3315 {
3316 if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_PIN))
3317 return -EINVAL;
3318 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
3319 }
3320 EXPORT_SYMBOL_GPL(pin_user_pages_fast);
3321
3322 /**
3323 * pin_user_pages_remote() - pin pages of a remote process
3324 *
3325 * @mm: mm_struct of target mm
3326 * @start: starting user address
3327 * @nr_pages: number of pages from start to pin
3328 * @gup_flags: flags modifying lookup behaviour
3329 * @pages: array that receives pointers to the pages pinned.
3330 * Should be at least nr_pages long.
3331 * @locked: pointer to lock flag indicating whether lock is held and
3332 * subsequently whether VM_FAULT_RETRY functionality can be
3333 * utilised. Lock must initially be held.
3334 *
3335 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
3336 * get_user_pages_remote() for documentation on the function arguments, because
3337 * the arguments here are identical.
3338 *
3339 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3340 * see Documentation/core-api/pin_user_pages.rst for details.
3341 *
3342 * Note that if a zero_page is amongst the returned pages, it will not have
3343 * pins in it and unpin_user_page*() will not remove pins from it.
3344 */
pin_user_pages_remote(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,int * locked)3345 long pin_user_pages_remote(struct mm_struct *mm,
3346 unsigned long start, unsigned long nr_pages,
3347 unsigned int gup_flags, struct page **pages,
3348 int *locked)
3349 {
3350 int local_locked = 1;
3351
3352 if (!is_valid_gup_args(pages, locked, &gup_flags,
3353 FOLL_PIN | FOLL_TOUCH | FOLL_REMOTE))
3354 return 0;
3355 return __gup_longterm_locked(mm, start, nr_pages, pages,
3356 locked ? locked : &local_locked,
3357 gup_flags);
3358 }
3359 EXPORT_SYMBOL(pin_user_pages_remote);
3360
3361 /**
3362 * pin_user_pages() - pin user pages in memory for use by other devices
3363 *
3364 * @start: starting user address
3365 * @nr_pages: number of pages from start to pin
3366 * @gup_flags: flags modifying lookup behaviour
3367 * @pages: array that receives pointers to the pages pinned.
3368 * Should be at least nr_pages long.
3369 *
3370 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
3371 * FOLL_PIN is set.
3372 *
3373 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3374 * see Documentation/core-api/pin_user_pages.rst for details.
3375 *
3376 * Note that if a zero_page is amongst the returned pages, it will not have
3377 * pins in it and unpin_user_page*() will not remove pins from it.
3378 */
pin_user_pages(unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages)3379 long pin_user_pages(unsigned long start, unsigned long nr_pages,
3380 unsigned int gup_flags, struct page **pages)
3381 {
3382 int locked = 1;
3383
3384 if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_PIN))
3385 return 0;
3386 return __gup_longterm_locked(current->mm, start, nr_pages,
3387 pages, &locked, gup_flags);
3388 }
3389 EXPORT_SYMBOL(pin_user_pages);
3390
3391 /*
3392 * pin_user_pages_unlocked() is the FOLL_PIN variant of
3393 * get_user_pages_unlocked(). Behavior is the same, except that this one sets
3394 * FOLL_PIN and rejects FOLL_GET.
3395 *
3396 * Note that if a zero_page is amongst the returned pages, it will not have
3397 * pins in it and unpin_user_page*() will not remove pins from it.
3398 */
pin_user_pages_unlocked(unsigned long start,unsigned long nr_pages,struct page ** pages,unsigned int gup_flags)3399 long pin_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
3400 struct page **pages, unsigned int gup_flags)
3401 {
3402 int locked = 0;
3403
3404 if (!is_valid_gup_args(pages, NULL, &gup_flags,
3405 FOLL_PIN | FOLL_TOUCH | FOLL_UNLOCKABLE))
3406 return 0;
3407
3408 return __gup_longterm_locked(current->mm, start, nr_pages, pages,
3409 &locked, gup_flags);
3410 }
3411 EXPORT_SYMBOL(pin_user_pages_unlocked);
3412